Healthcare Providers Verify

Healthcare Providers Verify

Validate practitioner credentials against the NPI registry and authoritative sources, powered by Nimble's web data APIs.

User request: $ARGUMENTS

Before running any commands, read references/nimble-playbook.md for Claude Code constraints (no shell state, no &/wait, sub-agent permissions, communication style).

---

Instructions

Step 0: Preflight + WSA Discovery

Sibling handoff check: Before running full preflight, check if healthcare-providers-extract or healthcare-providers-enrich ran earlier in this session by following the Sibling Handoff pattern from references/nimble-playbook.md. If same-day output exists, skip CLI check and profile load, and reuse WSA Layer 1/3 inventory. Only re-run Layer 2 if the verification focus changed.

Otherwise, run full preflight from references/nimble-playbook.md (5 simultaneous Bash calls: date calc, today, CLI check, profile load, index.md load).

Also simultaneously — run WSA discovery and setup:

• mkdir -p ~/.nimble/memory/{reports,healthcare-providers-verify/checkpoints}
• ls ~/.nimble/memory/healthcare-providers-verify/checkpoints/ 2>/dev/null
• Run Layer 1 (vertical) and Layer 3 (general tools) WSA discovery from

references/wsa-reference.md. Layer 2 (session-specific) runs after Step 1 when you know the user's specialty and verification focus.

Classify discovered agents into verification categories and validate with nimble agent get per references/wsa-reference.md.

From the preflight results:

• CLI missing or API key unset -> references/profile-and-onboarding.md, stop
• Tag all nimble CLI calls: nimble --client-source skill-healthcare-providers-verify <subcommand>. MCP path: not yet supported — see references/nimble-playbook.md for status.
• Profile exists -> note it for context. Determine mode using smart date windowing

from references/nimble-playbook.md:

• Full mode: first run OR last run > 14 days ago
• Quick refresh: last run < 14 days ago (re-verify only previously

Unverified/Flagged practitioners)

• Same-day repeat: if last_runs.healthcare-providers-verify is today, check

for existing report at ~/.nimble/memory/reports/healthcare-providers-verify-*[today].md. If found, ask: "Already ran today. Run again for fresh data?"

• No profile -> that's fine. This skill doesn't require onboarding. Proceed to Step 1.

Step 1: Parse Input + Starting Questions

Chained-from-sibling shortcut: Check for same-day extract or enrich output:

ls ~/.nimble/memory/reports/healthcare-providers-extract-*$(date +%Y-%m-%d).md 2>/dev/null
ls ~/.nimble/memory/reports/healthcare-providers-enrich-*$(date +%Y-%m-%d).md 2>/dev/null

If a same-day report exists, parse the {slug} and load the provider data (providers.json or enriched.json). This gives you names, credentials, specialties, and locations — skip parsing and go directly to Step 2.

Parse $ARGUMENTS for input type using the Input Parsing Pattern from references/nimble-playbook.md. Key routing:

• Sibling output detected (providers.json/enriched.json) -> proceed to Step 2
• CSV/Sheet/pasted data detected -> proceed to Step 2
• Unclear -> ask (counts as 1 of max 2 prompts)

If input is clear, confirm and ask one shaping question (plain text, not AskUserQuestion):

"Found N practitioners to verify. Quick questions:

1. What should I verify? (credentials, specialty, active status, practice address — or all)

2. Healthcare vertical? (ophthalmology, dental, dermatology, general, or other)"

If input is ambiguous, use AskUserQuestion (counts as 1 of max 2 prompts):

What practitioner data should I verify?

- Paste provider data directly (name + credentials + location, one per line)

- Provide a CSV file path or Google Sheet URL

- Or describe what you have (e.g., "a list of 50 ophthalmologists I need to

verify against the NPI registry")

Skip questions the user already answered in their initial message.

Step 2: Analyze Input Data

Parse the input into structured records. For each practitioner, identify:

• Claimed fields — name, credentials, specialty, state/city, practice name
• Verification targets — which claims to check based on user's focus

Minimum required fields: Name + at least one of (credentials, state, specialty). If a practitioner has only a name with no other identifiers, flag it: "Cannot verify [name] — need at least a state, credential, or specialty to search."

Build a verification plan summary:

"Analyzing N practitioners for verification:

- Names: N/N present

- Credentials claimed: N/N

- Specialty claimed: N/N

- State/location: N/N

>

Starting NPI verification..."

Run Layer 2 WSA discovery now that you know the specialty:

nimble agent list --limit 50 --search "[specialty]"
nimble agent list --limit 50 --search "[registry-user-mentioned]"

See references/wsa-reference.md for session-specific discovery.

Step 3: NPI Registry Lookup

Prefer the NPPES API — it returns structured JSON in one call instead of search + extract (two calls). Build the query URL from the practitioner's fields:

nimble extract --url "https://npiregistry.cms.hhs.gov/api/?version=2.1&first_name=[First]&last_name=[Last]&state=[ST]&limit=5" --format markdown

Add &taxonomy_description=[Specialty] if the specialty is known and specific enough. The API returns NPI number, status, credentials, taxonomy codes, addresses, and enumeration dates — everything needed for verification in a single call.

Fallback to web search if the NPPES API returns zero results or errors:

nimble search --query "[Name] [Credential] [State] NPI registry" --max-results 5 --search-depth lite

Then extract the top result from an NPI source (see source priority below).

Source priority (enforce in sub-agent prompts): 1. NPPES API (npiregistry.cms.hhs.gov/api/) — preferred, structured JSON 2. npidb.org/doctors/ — clean structured data, good fallback 3. nppes.cms.hhs.gov (provider-view pages) — official CMS source 4. Skip all others (healthline, hmedata, vitals, etc.) — inconsistent formatting

Tell sub-agents: "Only extract from NPPES API, npidb.org, or nppes.cms.hhs.gov. Ignore other NPI aggregator sites."

Search budget per provider: Max 3 search queries + 1 extraction per practitioner. If no NPI match after 3 attempts, mark as Unverified and move on. Tell sub-agents: "Do not run more than 4 nimble commands per provider. Mark as Unverified if no match by then."

For 10+ practitioners, use sub-agents (see Sub-Agent Strategy below).

Key fields from NPI records — see references/npi-verification-patterns.md for the full list: NPI number, status, credentials, taxonomy/specialty, enumeration date, last updated, practice address.

Checkpoint enforcement: After each sub-agent returns its batch results, the main context MUST write the checkpoint before spawning the next step or presenting results: 1. Receive sub-agent results 2. Write checkpoint: echo '[results]' > ~/.nimble/memory/healthcare-providers-verify/checkpoints/{slug}/batch-{n}.json 3. Continue to next step

Do NOT skip this — if the run fails between steps, the user loses all progress.

Step 4: Cross-Reference and Verify

For each practitioner, compare claimed data against extracted NPI data. Follow the verification logic in references/npi-verification-patterns.md:

1. Name matching — normalize both names and determine match level (Strong, Likely, Weak, No Match) per the name matching rules in the reference 2. Credential matching — compare claimed credentials against NPI record 3. Specialty matching — compare claimed specialty against NPI taxonomy using the taxonomy matching strategy in the reference 4. Address matching — compare claimed state/city against NPI practice address 5. Status check — verify NPI status is Active

Assign verification status per practitioner based on the totality of evidence:

• Verified — all claims match NPI record
• Partially Verified — NPI found, minor discrepancies
• Unverified — no NPI match found or unable to disambiguate
• Flagged — active mismatches requiring human review

See references/npi-verification-patterns.md for the detailed criteria for each status and the mismatch severity levels (Critical vs Warning).

Step 5: WSA Supplementary Verification (Optional)

If the user requested regulatory verification beyond NPI lookup, or if Step 5 left practitioners as Unverified that might benefit from additional sources:

Run verification-phase WSAs discovered in Step 0. See references/wsa-reference.md for the verification phase mapping, agent evaluation, and fallback chains.

Practice confirmation: For Unverified practitioners, try confirming their practice exists via practice-level WSAs or web search:

nimble search --query "[practice-name] [city] [state]" --max-results 5 --search-depth lite

Regulatory verification: For practitioners the user wants regulatory checks on:

nimble search --query "[name] [credentials] clinical trials OR FDA OR board certification" --max-results 5 --search-depth lite

Step 6: Deduplication & Confidence Scoring

Follow the Entity Deduplication pattern from references/nimble-playbook.md. Skill-specific dedup rules are in references/provider-extraction-patterns.md.

NPI dedup check: After all sub-agents return, scan for duplicate NPI numbers across batches. If two different providers mapped to the same NPI, flag both as "Flagged — possible NPI collision, requires human review." This catches data entry errors and name confusion in the source provider list.

Verification-specific scoring: The verification status (Verified / Partially Verified / Unverified / Flagged) replaces confidence scoring for this skill. Each status includes a confidence qualifier:

• High confidence — 2+ NPI sources corroborate, strong name match
• Medium confidence — single NPI source, likely name match
• Low confidence — weak name match, partial field matches

Step 7: Output

Present results as a verification report — showing status per practitioner with specific mismatch details. Group by verification status, include a "What This Means" section at the end.

# Provider Verification: [N] Practitioners Checked
*[Date] | [V] Verified, [PV] Partially Verified, [U] Unverified, [F] Flagged*

## TL;DR
Verified [V] of [T] practitioners against the NPI registry. [F] flagged for
review: [brief description of critical issues]. [U] could not be verified —
[common reason].

## Verification Results

| # | Name | Claimed | NPI Status | Verification | Issues | Source |
|---|------|---------|------------|-------------|--------|--------|
| 1 | Dr. Jane Smith | MD, Retinal Surgery, TX | Active (NPI 1234567890) | Verified | — | [NPI](url) |
| 2 | Dr. John Doe | OD, Ophthalmology, CA | Active (NPI 0987654321) | Partially Verified | Subspecialty differs | [NPI](url) |
| 3 | Dr. Alex Chen | MD, Dentistry, NY | Not Found | Unverified | No NPI match | [NPPES query](api-url) |
| 4 | Dr. Pat Lee | DO, Cardiology, FL | Deactivated | Flagged | NPI deactivated 2024-01 | [NPI](url) |

## Flagged Practitioners (Requires Human Review)

### Dr. Pat Lee
**Claimed:** DO, Cardiology, FL
**NPI Record:** NPI 1122334455 — **Deactivated** (01/15/2024)
**Issues:**
- CRITICAL: NPI status is Deactivated since January 2024
- Credential matches (DO confirmed)
- Specialty matches (Cardiovascular Disease taxonomy)
**Source:** [NPI Record](url)
**Action needed:** Confirm if provider has re-registered or if this is a
different individual.

[Repeat per flagged practitioner]

## Unverified Practitioners

[List practitioners where no NPI match was found, with search queries attempted]

## Verification Summary
- **Verified:** [V] practitioners — all claims confirmed
- **Partially Verified:** [PV] — minor discrepancies noted
- **Unverified:** [U] — no NPI match (common names, missing identifiers)
- **Flagged:** [F] — critical issues requiring review

## Sources
[Clickable URL for every NPI lookup page used, grouped by practitioner]

## What This Means
[Actionable interpretation: which practitioners are safe to include in your
directory, which need follow-up, what the verification rate tells you about
your data quality. Suggest next steps for unverified/flagged records.]

Source links are mandatory. Every verification finding must trace back to a source URL.

Step 8: Save to Memory

Make all Write calls simultaneously:

• Report -> ~/.nimble/memory/reports/healthcare-providers-verify-{slug}-{date}.md
• Verification data -> ~/.nimble/memory/healthcare-providers-verify/{slug}/verified.json
• Profile -> update last_runs.healthcare-providers-verify in

~/.nimble/business-profile.json (only if profile exists)

• Follow the wiki update pattern from references/memory-and-distribution.md: update

index.md rows for all affected entity files, append a log.md entry for this run.

• Clean up checkpoint (complete run) or keep (partial run)

Update sibling artifacts: If providers.json or enriched.json exists for this slug under ~/.nimble/memory/, merge NPI numbers and verification status into those files. Generate a verified CSV export at ~/.nimble/memory/healthcare-providers-verify/{slug}/verified-{date}.csv with all verification columns (NPI, NPI Status, NPI Taxonomy, Verification Status). Offer this export path in Step 9 so the user can copy it where needed.

Step 9: Share, Distribute & Follow-ups

Always offer distribution — do not skip. Follow references/memory-and-distribution.md for connector detection and sharing flow.

Notion: full verification report as a dated subpage. Slack: TL;DR with verification counts and flagged items only.

Follow-ups:

• "Tell me more about Dr. X" -> show full verification detail
• "Export as CSV" -> generate CSV with verification statuses
• "Re-verify flagged only" -> re-run NPI search for Flagged/Unverified only
• "What should I do about the flagged ones?" -> actionable next steps per issue

Sibling skill suggestions:

Next steps:

- Run healthcare-providers-extract on unverified providers' practice URLs

to get fresh data

- Run healthcare-providers-enrich to fill gaps in verified providers' records

- Run market-finder to find additional practices in this area

---

Sub-Agent Strategy

For batch verification (10+ practitioners), use nimble-researcher agents (agents/nimble-researcher.md) to parallelize NPI lookups and extraction.

Follow the sub-agent spawning rules from references/nimble-playbook.md (bypassPermissions, batch max 4, explicit Bash instruction, fallback on failure).

Spawn pattern: One agent per batch of 5 practitioners. Each agent runs Steps 3-4 for its assigned practitioners and returns verification records. Tell each agent to use nimble extract-batch for its NPI result URLs where possible — one batch call per agent is faster than sequential calls.

Small batch optimization: If fewer than 10 practitioners, run directly from the main context instead of spawning agents.

Fallback: If any agent fails, run those verifications directly from the main context. Never leave gaps in the output.

---

Error Handling

See references/nimble-playbook.md for the standard error table (missing API key, 429, 401, empty results, extraction garbage). Skill-specific errors:

• No NPI results for practitioner: "Couldn't find an NPI record for [name] in

[state]. The name may be too common, or the provider may practice under a different name. Want me to try with additional context (practice name, NPI number)?"

• Multiple NPI matches: "Found multiple NPI records for [name] in [state].

Can you confirm which one? [list top 3 with NPI numbers and specialties]"

• NPI page extraction returned garbage: "The NPI lookup page appears to be

JavaScript-rendered. Retrying with browser rendering..." (auto-retry with --render per the shared pattern)

• CSV/Sheet parse error: "Couldn't parse the input file. Expected columns with

practitioner names and at least one identifier (state, specialty, or credentials). Can you paste the data directly instead?"

• Insufficient data for verification: "Cannot verify [N] practitioners — they

have only a name with no state, credential, or specialty. Add identifiers or remove them from the list."

Memory & Distribution

How skills persist knowledge across sessions and distribute reports to external tools.

---

Memory Architecture

All persistence lives under ~/.nimble/ — never touch user project files.

~/.nimble/
├── business-profile.json          # Tier 1: Hot cache (see profile-and-onboarding.md)
└── memory/                        # Tier 2: Deep storage (loaded on demand)
    ├── index.md                   # Global index (one line per directory)
    ├── log.md                     # Chronological activity log (append-only)
    ├── backlog.md                 # Research questions and knowledge gaps
    ├── synthesis/                 # Cross-entity analysis pages
    ├── competitors/               # Accumulated intel per competitor
    │   └── index.md              # Per-directory entity catalog
    ├── people/                    # Contact profiles for meeting prep
    │   └── index.md
    ├── companies/                 # Deep-dive research results
    │   └── index.md
    ├── reports/                   # Timestamped full skill outputs
    ├── positioning/               # Per-competitor positioning snapshots
    │   └── index.md
    └── glossary.md                # Industry terms and jargon

Tier 1 (business-profile.json) — loaded every session. See references/profile-and-onboarding.md for the full schema and update patterns.

Tier 2 (memory/) — loaded on demand when a skill needs deeper context.

Wiki Primitives

The memory directory includes wiki-level files that make the knowledge base navigable, queryable, and self-maintaining:

~/.nimble/memory/
├── index.md                       # Global summary (one line per directory)
├── log.md                         # Chronological activity log (append-only)
├── backlog.md                     # Research questions and knowledge gaps
├── synthesis/                     # Cross-entity analysis pages
│   ├── index.md                   # Per-directory catalog (same format as others)
│   └── competitive-landscape.md   # (created dynamically when patterns emerge)
├── competitors/
│   ├── index.md                   # Per-directory entity catalog
│   ├── widgetco.md
│   └── gizmotech.md
├── people/
│   ├── index.md
│   └── alex-kim.md
├── companies/
│   ├── index.md
│   └── ...
├── reports/
├── positioning/
│   ├── index.md
│   └── ...
└── glossary.md

Skills create index files, log.md, and synthesis/ on first write if missing.

---

Wiki Content Index (Two-Tier)

Indexes live at two levels: a lightweight global index for cross-directory navigation, and per-directory indexes for detailed entity catalogs.

Global Index (~/.nimble/memory/index.md)

One line per directory — entity count and last-updated date. Never lists individual entities. Stays under 30 lines forever.

# Knowledge Index

| Directory | Entities | Last Updated |
|-----------|----------|-------------|
| [[competitors/index]] | 5 | 2026-03-20 |
| [[people/index]] | 3 | 2026-03-15 |
| [[companies/index]] | 8 | 2026-03-18 |
| [[positioning/index]] | 5 | 2026-03-20 |
| [[synthesis/index]] | 2 | 2026-03-20 |

Per-Directory Index ({dir}/index.md)

One row per entity file with summary and last-updated date. Owned by the skills that write to that directory. Scales independently — each directory can grow without affecting other indexes.

# Competitors Index

| File | Summary | Updated |
|------|---------|---------|
| [[competitors/widgetco]] | Enterprise SaaS competitor, Series C | 2026-03-20 |
| [[competitors/gizmotech]] | API-first competitor, growing fast | 2026-03-20 |

Rules

• Skills read only their directory's index in preflight. competitor-intel reads

competitors/index.md; meeting-prep reads people/index.md. Cross-directory lookups go through the global index first, then the relevant directory index.

• Skills update only their directory's index on write. When a skill creates or

updates an entity file, update the row in that directory's index. Use the entity file's first # Heading as the summary if none exists.

• Global index is updated after directory index changes. Bump the entity count

and last-updated date for the affected directory.

• Created on first skill run if missing. Skills should not fail if an index

doesn't exist — create it with whatever entities are written in that run.

• Obsidian-compatible. [[path/entity]] links (without .md extension) work as

wiki links in Obsidian. Path is relative to ~/.nimble/memory/.

---

Chronological Wiki Log (log.md)

~/.nimble/memory/log.md is an append-only timestamped record of skill runs and findings. Grep-friendly format for answering "what did I learn this week?"

# Activity Log

## [2026-03-15] meeting-prep
- Updated: [[people/alex-kim]], [[companies/widgetco]]
- Key findings:
  - Alex Kim moved to VP Engineering role
  - Interested in API performance benchmarks

## [2026-03-18] company-deep-dive
- Created: [[companies/target-corp]]
- Key findings:
  - Series B closed at $30M, Sep 2025
  - Expanding into EU market Q2 2026

## [2026-03-20] competitor-intel
- Created: [[competitors/widgetco]], [[competitors/gizmotech]]
- Updated: [[competitors/acme-rival]]
- Key findings:
  - WidgetCo launched enterprise tier pricing
  - GizmoTech hired new CTO from CloudCorp

Rules

• Append at the end of the file (oldest first, newest last). Normal writes are

pure appends — no read-insert-rewrite needed. LLMs read the whole file; humans use grep "^## \[" log.md | tail -10 for recent entries.

• Format: ## [YYYY-MM-DD] skill-name — enables grep "^## \[" log.md | tail -10.
• Content: List entities created/updated (as [[path/entity]] links), then 2-3

bullet points of key findings. Keep entries concise — this is a log, not a report.

• Rotate entries older than 90 days as a separate maintenance step. After

appending the new entry, check the oldest entries (at the top). If older than 90 days, remove them. This rotation is not part of the normal append — it's a periodic cleanup that triggers during writes. The full reports in reports/ are the permanent record; log.md is for recent activity scanning.

• Created on first skill run if missing.

---

Cross-Entity References

Entity files use Obsidian-compatible [[path/entity]] wiki links to connect related entities across directories.

Format

# Alex Kim

## Current Role
VP of Engineering at [[competitors/widgetco]] (since 2024)

## Related
- Employer: [[competitors/widgetco]]
- Previous: [[companies/cloudcorp]]

# WidgetCo

## Key People
- [[people/alex-kim]] — VP Engineering
- [[people/jane-smith]] — CEO

## Related Competitors
- [[competitors/gizmotech]] — overlapping market segment

Rules

• Link format: [[directory/entity-slug]] — no .md extension, path relative

to ~/.nimble/memory/. Obsidian resolves these as wiki links.

• Add cross-references when relationships are discovered. When a skill finds

that a person works at a tracked company, or two competitors share a market segment, add links in both directions.

• Handle missing targets gracefully. A cross-reference to a file that doesn't

exist yet is fine — it becomes a valid link once that entity is created. Skills should not fail on dangling links.

• Skills follow links to enrich output. When meeting-prep finds

[[competitors/widgetco]] in a person's file, it loads that competitor file for additional context. When competitor-intel finds [[people/alex-kim]] in a competitor file, it can surface that relationship in the briefing.

When to Add Cross-References

---

Ad-Hoc Insights

When a user signals "save this", "remember that", "note this down", or similar intent during a conversation, file the insight into the relevant entity file(s) instead of letting it vanish into chat history.

Filing Pattern

1. Identify the relevant entity file(s). If the insight is about a competitor, file it in competitors/{name}.md. If it spans multiple entities (e.g., "WidgetCo is partnering with GizmoTech"), update all relevant files. 2. Append under a dated `## Insights` section:

## Insights
### 2026-03-22
- User noted: WidgetCo's enterprise pricing is 2x ours — [[competitors/gizmotech]]
  is closer to our price point [ad-hoc]

3. Add cross-references if the insight connects entities (as shown above). 4. Update the directory's `index.md` — bump the last-updated date for the affected file(s), and update the global index.md counts. 5. Append to `log.md` (at the end of the file):

## [2026-03-22] ad-hoc-insight
- Updated: [[competitors/widgetco]], [[competitors/gizmotech]]
- Key findings:
  - WidgetCo enterprise pricing is 2x user's, GizmoTech closer to parity

Rules

• Tag with `[ad-hoc]` so skills can distinguish user-contributed insights from

skill-generated findings during dedup.

• Multi-entity insights update all relevant files with cross-references between

them.

• Don't create entity files for throwaway comments. If the user says "remember

that meetings on Fridays are bad", that's a preference (update business-profile.json), not an entity insight.

---

Cross-Entity Synthesis Pages

~/.nimble/memory/synthesis/ contains pages that analyze patterns across multiple entity files. Unlike entity files (which accumulate facts about one entity), synthesis pages draw conclusions across the knowledge base.

Page Creation

Synthesis pages are created dynamically when patterns emerge across entities — not from a pre-defined list. Common examples:

Page names should be slug-formatted topic labels (not skill names). Skills create synthesis pages when a pattern recurs across 3+ entities — this keeps synthesis data-driven rather than speculative.

Format

Synthesis pages use YAML frontmatter to track which entity files they were built from and when. This makes staleness deterministic — compare current file timestamps against the recorded ones.

---
confidence: high
sources:
  - path: competitors/widgetco.md
    updated: 2026-03-20
  - path: competitors/gizmotech.md
    updated: 2026-03-20
  - path: competitors/acme-rival.md
    updated: 2026-03-18
generated_by: competitor-intel
generated_at: 2026-03-20
---
# Competitive Landscape

## Market Map
[Positioning of each competitor by segment, size, strategy]

## Feature Comparison
| Capability | Us | [[competitors/widgetco]] | [[competitors/gizmotech]] |
|---|---|---|---|
| Real-time data | ✅ | ❌ | Partial |

## Pricing Comparison
[Tier-by-tier comparison where known]

## Key Patterns
- Trend 1 with evidence from multiple competitors
- Trend 2 with cross-entity citations

## What This Means
[Strategic implications — what the patterns suggest for the user's company]

Rules

• Cite source entity files with [[path/entity]] links. Every claim must trace

back to an entity file.

• Track sources and confidence in frontmatter. confidence: high|medium|low

reflects data completeness (high = all key sources available, low = sparse data). The sources: block lists every entity file used and its last-modified date at generation time. To check staleness, compare current file timestamps against the recorded ones — if any source was updated since generation, the page is stale.

• Refresh when sources are stale. If a skill adds major new signals to 2+ source

entities since the synthesis was generated, regenerate. Don't regenerate on every run — only when the source timestamps diverge.

• Use `nimble-analyst` agent for synthesis generation. The analyst has the right

model (Sonnet) for cross-entity pattern recognition and strategic analysis.

Generation Trigger

competitor-intel generates competitive-landscape.md when:

• 3+ competitors have been researched in the current run, OR
• The existing synthesis page's source timestamps are stale (source entities were

updated since generation)

Other synthesis pages are created by the relevant skills when patterns emerge, or on user request.

---

Research Backlog (backlog.md)

~/.nimble/memory/backlog.md tracks knowledge gaps and research questions — things to investigate in future skill runs. This is not synthesis (derived, read-only output) — it's imperative (drives future action).

# Research Backlog

## Open
- [ ] WidgetCo pricing for enterprise tier — couldn't find public pricing [2026-03-20, competitor-intel]
- [ ] GizmoTech Series B details — rumored but unconfirmed [2026-03-20, competitor-intel]
- [ ] Alex Kim's LinkedIn activity — profile was private [2026-03-15, meeting-prep]

## Resolved
- [x] WidgetCo new CTO name — confirmed: Sarah Chen [2026-03-22, competitor-intel]

Rules

• Any skill can append questions to the ## Open section when it encounters

gaps during research. Tag each with date and skill name.

• Users can add questions via ad-hoc insights ("find out about X next time").
• Skills check backlog before running to avoid re-researching resolved questions

and to prioritize open ones relevant to the current run.

• Resolved questions get moved to ## Resolved with a resolution date — not

deleted. This preserves the audit trail.

---

Deep Storage Formats

competitors/

One file per competitor. Append new findings under dated headers — never overwrite.

# WidgetCo

## Key Facts
- Domain: widgetco.com
- HQ: San Francisco
- Funding: Series C ($45M, Jan 2026)
- CEO: Jane Smith

## Signals
### 2026-03-20
- Launched new enterprise tier pricing — [source URL]
- Hired VP of Sales from CRMHub — [source URL]

### 2026-03-13
- Announced partnership with AWS — [source URL]

people/

One file per contact. Used by meeting-prep skill.

# Alex Kim

## Current Role
VP of Engineering at WidgetCo (since 2024)

## Background
- Previously: Senior Director at CloudCorp (2019-2024)
- Education: MS Computer Science, top-10 program

## Notes from Previous Meetings
### 2026-03-15
- Interested in our API performance benchmarks
- Prefers technical depth over high-level summaries

companies/

Detailed company profiles from deep-dive research.

# Target Corp

## Overview
- Industry: Enterprise SaaS | Founded: 2015 | HQ: Austin, TX | ~500 employees

## Financials
- Last funding: Series B ($30M, Sep 2025) | Revenue: Est. $40M ARR

## Recent News
(dated entries, same format as competitors/)

reports/

Timestamped full skill outputs. Save the complete briefing, not a summary.

Naming: {skill-name}-{YYYY-MM-DD}.md — if a skill may produce multiple reports per day (e.g., meeting-prep for different companies), add a qualifier: {skill-name}-{qualifier}-{YYYY-MM-DD}.md. The qualifier is defined in each skill's SKILL.md (e.g., company slug for meeting-prep).

glossary.md

Industry terms and jargon. Updated when the user uses unfamiliar terms.

Bootstrapping (First Run)

mkdir -p ~/.nimble/memory/{competitors,people,companies,reports,positioning,synthesis}

Create stub files for each competitor from the onboarding flow.

index.md and log.md are created automatically on the first skill run that writes to memory — no need to create empty stubs during bootstrapping.

Differential Analysis

The key feature across all skills — only surface what's genuinely new.

Dedup Lifecycle

Memory loading happens at two points in every skill:

1. Step 0 (Preflight): Load relevant memory files for context. This tells the skill what's already known so it can pass known signals to sub-agents for dedup during research. For example, competitor-intel loads ~/.nimble/memory/competitors/*.md; meeting-prep loads ~/.nimble/memory/people/*.md.

2. Analysis step (before report generation): Final dedup check. Compare all findings from research against loaded memory. Only signals classified as NEW or UPDATED (per the freshness classification in nimble-playbook.md) make it into the report.

What "new" means

• "WidgetCo raised a Series C" is noise if already in memory
• "WidgetCo just hired a new CTO" is a new signal worth highlighting
• "WidgetCo raised a Series C" with a new detail (amount, lead investor) is an UPDATE

Learning from Corrections

When the user corrects the skill, update both tiers:

Always confirm the update to the user.

Checkpointing & Resume

For multi-phase pipelines (map → extract → enrich → score), save intermediate results so failed or interrupted runs can resume without re-doing completed work.

Storage

~/.nimble/memory/{skill-name}/checkpoints/{slug}/
├── map.json           # Phase 1 output
├── extract.json       # Phase 2 output
└── enrich.json        # Phase 3 output

{slug} is a stable identifier derived from the run's input parameters (e.g., URL domain, search query hash). Same input = same slug = resumable.

Checkpoint format

Each phase file is JSON:

{
  "phase": "extract",
  "status": "complete",
  "timestamp": "2026-04-03T15:30:00Z",
  "record_count": 47,
  "data": [ ... ]
}

status is "complete" or "partial" (interrupted mid-phase).

Resume logic

On re-run with the same parameters: 1. Detect existing checkpoint directory for the slug 2. Offer: "Found previous run (47 records from Apr 3). Resume and fill gaps, or start fresh?" 3. If resume: skip phases where status = "complete", re-run where status = "partial" or file is missing 4. If start fresh: delete the checkpoint directory and begin from phase 1

Rules

• One checkpoint directory per unique run (keyed by slug)
• Clean up checkpoints older than 30 days on skill startup
• Don't checkpoint trivial runs (< 5 records) — the overhead isn't worth it

Rules

• Never touch user project files. All persistence under ~/.nimble/.
• Append, don't overwrite. Deep storage grows over time with dated sections.
• Read on demand. Only load files when the skill actually needs them.
• Update profile after every run. At minimum, last_runs timestamp.
• Update wiki files after every memory write. Update the directory's index.md

for affected entities, bump the global index.md counts, append a log.md entry for the run, and add cross-references where relationships are discovered.

• Handle missing gracefully. If a file doesn't exist, create it. This includes

index files, log.md, backlog.md, and cross-reference targets.

---

Source Links Enforcement

Every signal in every report must include a clickable source URL. This is a hard requirement — reports without source links are incomplete and must not be distributed.

What counts as a source link:

• A direct URL to the article, press release, or page where the signal was found
• The URL returned by nimble search in the result's url field
• For extracted content, the URL passed to nimble extract --url

What does NOT count:

• A company's homepage (unless the signal is specifically about homepage content)
• A generic domain without a path (e.g., https://widgetco.com)
• "Source: web search" or any non-clickable attribution

If a signal has no source URL after research and extraction, drop it from the report. An unsourced signal is worse than a missing one — it can't be verified and erodes trust.

---

Report Distribution

After presenting output, offer sharing based on available MCP connectors.

Connector Detection

Check before presenting options:

• Notion: mcp__plugin_Notion_notion__notion-create-pages
• Slack: Any Slack MCP tool

Sharing Flow

Use AskUserQuestion with only the available options:

Share this report?

- Save to Notion — full report as a page

- Send to Slack — TL;DR to a channel

- Both

- Skip

Notion: Create a dated subpage. If integrations.notion.reports_page_id exists in the profile, use it as parent. Otherwise ask and save the ID for next time.

Slack: Post TL;DR only — Slack is for alerts, not full reports. If integrations.slack.channel exists, use it. Otherwise ask and save.

Neither available (first run only):

Tip: If you connect a Notion or Slack MCP server, I can save reports or post

TL;DRs to your team automatically.

Don't repeat this tip on subsequent runs.

Nimble Playbook

How to run Nimble CLI commands in Claude Code. Read this before executing any commands.

---

Claude Code Execution Rules

• No shell state persistence. Variables set in one Bash call are gone in the next.

Inline all values (dates, paths, names) directly into every command.

• No `&` + `wait` parallelism. It breaks in Claude Code. Instead, make **multiple

Bash tool calls in a single response** — they run in parallel natively.

• Search returns JSON — --output-format doesn't change this. With `--search-depth

lite`, the JSON is small (title, description, URL per result). Parse it directly.

• Extract returns JSON with `data.markdown` — use --format markdown to get clean

content in the data.markdown field.

Preflight Pattern

Transport selection (run once per session)

Skills work via two transports — CLI (preferred, full surface area) or MCP (fallback, curated tool set covering the same operations). Pick one at the start of every session and stick with it; don't re-probe on every command.

Connector not connected (Cowork / claude.ai) — verify BEFORE working

In Cowork / claude.ai the plugin is often installed while its connector is not yet connected, so live data calls fail. Confirming the connection is a required preflight step — not an error to react to mid-task. When mcp__plugin_nimble_nimble__* tools are listed but you haven't confirmed the connector is live, run one read-only probe before any real work:

• A single nimble_agents_list call is the cheapest confirmation. Success →

connected, proceed. Auth / not-connected error, or a response containing an OAuth authorization URL → not connected.

When not connected, surface this verbatim and stop — do not fall back to WebFetch, WebSearch, curl, or any other tool, and do not guess at data:

Your Nimble plugin is installed, but its connector isn't connected yet — that's

why I can't fetch live data. To connect it:

>

1. Open Customize → Connectors

2. Find Nimble and click Connect

3. Complete the login in your browser. No Nimble account? You can create one

right there during login.

4. Once it shows Connected, re-run your request and I'll continue.

If a tool hands back an OAuth "Authorize" URL

A not-connected tool call may return an authorization link (e.g. "Authorize Nimble MCP →") instead of data. Present that link to the user exactly as given, then stop and wait. Hard rules:

• Never invent a completion flow. There is no "paste the URL from your address

bar back to me" step, and you cannot "complete the connection" yourself. Claiming either is a hallucination.

• Never say the tools "will activate" and then call them in the same turn. Wait

for the user to confirm they've authorized, then retry.

• To check whether authorization succeeded, run one read-only nimble_agents_list

probe — don't assume.

No plugin and no CLI

If neither path works at all (no plugin installed, no CLI installed), surface this hint verbatim and stop:

Nimble isn't installed. Pick the path for your environment:

>

Any Claude product (Claude Code, Claude Cowork, claude.ai) — recommended:

```

/plugin install nimble

```

Installs the Nimble plugin. The .mcp.json inside the plugin auto-registers as a Connector in Customize → Connectors. First tool call triggers the OAuth flow — no API key needed.

>

Codex CLI or other terminal agents (shell access, no `/plugin`):

```

npm i -g @nimble-way/nimble-cli

```

Then export NIMBLE_API_KEY=<key> and re-run. See references/profile-and-onboarding.md for the full install flow.

>

Cursor, VS Code, or any other MCP client:

Paste this into your MCP settings (.cursor/mcp.json or host equivalent):

```json

{

"mcpServers": {

"nimble": { "type": "http", "url": "https://mcp.nimbleway.com/mcp" }

}

}

```

The plugin path (/plugin install nimble) is the easiest onboarding everywhere it works — one command, OAuth handles auth, no API key to manage. Use the CLI path only when shell access is available but /plugin install isn't (Codex, raw terminal agents). Use the manual mcp.json path only for MCP clients outside the Claude family.

Standard preflight (run in parallel after transport is selected)

Every skill kicks off with these simultaneous calls:

• python3 -c "from datetime import datetime, timedelta; print((datetime.now() - timedelta(days=14)).strftime('%Y-%m-%d'))" (14 days ago)
• date +%Y-%m-%d (today)
• cat ~/.nimble/business-profile.json 2>/dev/null (profile — fall back to MCP filesystem tool if shell unavailable)
• cat ~/.nimble/memory/index.md 2>/dev/null (global wiki index — know what directories have data)

Don't skip the transport check — running CLI commands when only MCP is available (or vice versa) wastes a turn and confuses the user.

Request Attribution

All Nimble API calls must carry a client_source tag so usage can be tracked per skill. The value is always skill- followed by the exact SKILL.md name field (e.g. skill-competitor-intel, skill-seo-intel, skill-nimble-web-expert).

CLI path — add --client-source skill-{name} as the global flag on every nimble command. Place it immediately after nimble, before the subcommand. No shell state persistence means this must be inlined on every individual call:

nimble --client-source skill-{name} search --query "..."
nimble --client-source skill-{name} extract --url "..."
nimble --client-source skill-{name} agent run --agent <name> --params '{...}'
nimble --client-source skill-{name} map --url "..."
nimble --client-source skill-{name} crawl run --url "..."

MCP path — per-skill client source tracking is not yet supported by the MCP server (it currently sends X-Client-Source: nimble_mcp_server for all calls regardless of skill). This will be enabled once the MCP server adds CLIENT_SOURCE support — no action needed here until then.

Sibling Handoff

When skills in the same family chain together (e.g., extract → enrich → verify), the second skill can skip redundant preflight work. Detect a sibling handoff by checking for same-day output from the upstream skill:

ls ~/.nimble/memory/reports/{upstream-skill}-*$(date +%Y-%m-%d).md 2>/dev/null

Use the dated report as the recency signal — data files under memory/{skill}/ may not have dates in their filenames, so always verify via the report timestamp. If a same-day report exists, parse the slug from the filename and load the corresponding data files.

If same-day sibling output exists:

• Skip CLI check and profile load — they were validated minutes ago
• Reuse WSA Layer 1 and Layer 3 inventory — the catalog hasn't changed. Only

re-run Layer 2 if the specialty or context changed.

• Use the sibling's structured output directly — if the upstream skill produced

data files with domains and page URLs, don't re-search for what's already known. Construct URLs from known patterns instead of running N web searches.

If no same-day sibling output exists: Run full preflight as normal.

This pattern is optional — skills MUST still work standalone without sibling output. The handoff is a fast path, not a requirement.

Smart Date Windowing

For any skill using --start-date based on previous runs:

• First run: 14 days ago → full mode
• Last run < 3 days ago: use 7 days ago (too narrow = empty results) → quick refresh
• Last run 3-14 days ago: use the last run date → quick refresh
• Last run > 14 days ago: 14 days ago → full mode
• Same-day repeat: if last_runs.{skill-name} is today, check if a report already

exists at ~/.nimble/memory/reports/{skill-name}*[today].md. If it does, ask the user before re-running: "Already ran today. Run again for fresh data?" Don't silently re-run — it wastes API credits and produces near-identical output. Exception — meeting-prep: Skip the same-day report check. Meeting-prep is per-meeting, not per-day — users may prep for multiple meetings in a single day. Instead, meeting-prep checks freshness at the entity level: load cached profiles from ~/.nimble/memory/people/ and ~/.nimble/memory/companies/ and offer to reuse recent research rather than blocking the run.

---

Search

# Standard search (always use --search-depth lite for discovery)
nimble search --query "company name news" --max-results 10 --search-depth lite

# News-focused search
nimble search --query "company name" --focus news --max-results 10 --search-depth lite

# Date-filtered search (inline the date — don't use variables)
nimble search --query "company funding" --focus news --start-date "2026-03-11" --max-results 10 --search-depth lite

# Social signals from X/LinkedIn
nimble search --query "Company" --include-domain '["x.com", "linkedin.com"]' --max-results 10 --search-depth lite --time-range week

# Deep search (full page content — only for comprehensive analysis, costs more)
nimble search --query "company name" --search-depth deep --max-results 5

# Fast search (premium tier — not used by default)
# nimble search --query "company name" --search-depth fast --max-results 10

Key flags:

• --query — search query string (required)
• --focus — general, news, shopping, social, coding, academic.

`social` searches social platform people indices directly (LinkedIn, X) — best for finding specific people. If it errors, use --include-domain '["linkedin.com"]' as an alternative approach.

• --max-results — max results to return
• --start-date / --end-date — date filters (YYYY-MM-DD)
• --search-depth — lite (1 credit), deep (1 + 1/page)
• --include-domain — JSON array of domains, e.g., '["x.com", "linkedin.com"]'
• --time-range — e.g., week
• --country — geo-targeted results (e.g., "US", "IL")
• --include-answer — LLM-powered answer summary

Date range strategy:

• First run: 14 days ago
• Subsequent runs: last_runs timestamp from business profile
• If < 3 results: retry without --start-date

Extract

# Extract article content as markdown (default for content analysis)
nimble extract --url "https://example.com/article" --format markdown

# Extract raw HTML (required for <head> metadata: canonical, schema, og, meta tags)
nimble extract --url "https://example.com" --format html

# Extract with JavaScript rendering (for dynamic/SPA pages)
nimble extract --url "https://example.com/spa" --render --format markdown

Response is JSON. The field returned depends on --format:

• --format markdown → data.markdown (clean body content)
• --format html → data.html (raw HTML including <head>)
• --format plain_text → data.plain_text
• --format simplified_html → data.simplified_html

Format selection by use case:

Key flags:

• --url — target URL (required)
• --format — markdown, html, simplified_html, plain_text (pick based on table above)
• --render — render JavaScript using a browser
• --parse --parser '{...}' — structured extraction via LLM parser schema

Extraction fallback (if data.markdown is mostly JavaScript/boilerplate): 1. Garbage check: If data.markdown has < 100 characters of meaningful content (after stripping nav/footer boilerplate), treat it as garbage. 2. Retry with --render --format markdown (handles JS-heavy/SPA pages) 3. If still garbage: search for the same article title on a different domain 4. If still nothing: skip and log — never abort a batch for a single extraction failure

Extract async & batch

# Async — submit single URL, get task_id, poll for results
nimble extract-async --url "https://example.com/page" --render --format markdown

# Batch — up to 1,000 URLs in one request
nimble extract-batch \
  --shared-inputs 'render: true' --shared-inputs 'format: markdown' \
  --input '{"url": "https://example.com/page-1"}' \
  --input '{"url": "https://example.com/page-2"}'

Poll async tasks with nimble tasks get --task-id <id> and fetch results with nimble tasks results --task-id <id>. Poll batches with nimble batches progress --batch-id <id>.

Map & Site Mapping

nimble map --url "https://example.com/blog" --limit 20

Site Mapping Pattern

Use nimble map to discover a site's page structure, then score and filter pages by relevance before extracting.

1. Discover: nimble map --url {url} --limit {cap} — returns a list of URLs 2. Score: Each skill defines a keyword/weight table for URL path segments (e.g., /providers = High, /about = Medium, /blog = Low). Score each discovered page against the table. 3. Filter: Keep pages scoring above the skill's threshold. Always include the homepage as a fallback. 4. Fallback: If nimble map returns < 3 candidates, use nimble search --query "site:{domain} {keywords}" --max-results 10 --search-depth lite

Each skill provides its own keyword/weight table in SKILL.md — the pattern here is the discover → score → filter → fallback flow.

Agents

Pre-built extraction templates for structured data from specific sites (Amazon, LinkedIn, Google, etc.). Use when you need structured fields rather than raw page content.

# List available agents (search by domain or vertical)
nimble agent list --limit 100
nimble agent list --limit 100 --search "amazon"

# Inspect an agent's schema (input params + output fields)
nimble agent get --template-name <agent_name>

# Run an agent (sync — waits for result)
nimble agent run --agent <agent_name> --params '{"key": "value"}'

# Run an agent (async — returns task_id, poll for results)
nimble agent run-async --agent <agent_name> --params '{"key": "value"}' \
  --callback-url "https://your.server/callback"

Key flags for `run` / `run-async`:

• --agent — agent name from nimble agent list (required)
• --params — JSON object with agent input parameters (required)
• --localization — enable zip_code/store_id localization (agent-dependent)

Additional flags for `run-async`:

• --callback-url — POST callback when task completes
• --storage-type — s3 or gs
• --storage-url — destination bucket URL
• --storage-compress — gzip the stored output
• --storage-object-name — custom filename instead of task_id

Response: data.parsing contains the structured output. Shape depends on agent type:

• PDP (product/profile/detail) → flat dict
• SERP / list → array of objects
• Google Search → {"entities": {"OrganicResult": [...], ...}}

Async task states: pending → success or error. Poll with nimble tasks results --task-id <task_id>.

Fallback rule: If no agent exists for the target domain, fall back to nimble search + nimble extract. Don't fail silently — log which domains lacked agent coverage so agent-builder can fill gaps later.

Agent batch

# Up to 1,000 agent requests in one call
nimble agent run-batch \
  --shared-inputs 'agent: amazon_serp' \
  --input '{"params": {"keyword": "iphone 15"}}' \
  --input '{"params": {"keyword": "iphone 16"}}'

Returns a batch_id. Poll with nimble batches progress --batch-id <id>, then fetch individual results with nimble tasks results --task-id <id>.

Tasks & batches polling

# Single async task
nimble tasks get --task-id <task_id>          # check status
nimble tasks results --task-id <task_id>      # fetch results

# Batch
nimble batches progress --batch-id <batch_id> # lightweight progress check
nimble batches get --batch-id <batch_id>      # all task IDs + states
nimble batches list --limit 20                # list all batches
nimble tasks list --limit 20                  # list all tasks

Workflow: Always nimble agent get before nimble agent run to understand the expected input params and output fields.

Agent Creation (generate → poll → iterate → publish)

Create custom extraction agents for any website. The full lifecycle is available via CLI.

# Generate a new agent
nimble agent generate \
  --agent-name niche_store_pdp \
  --prompt "Extract product name, price, rating, and first 5 reviews" \
  --url "https://example.com/products/widget-pro"

# Refine an existing agent (clone + apply new prompt)
nimble agent generate \
  --agent-name niche_store_pdp \
  --from-agent niche_store_pdp \
  --prompt "Add a discount_percentage field"

# Poll generation status (async — typically 1-3 min)
nimble agent get-generation --generation-id <generation_id>

# Publish when satisfied
nimble agent publish --agent-name niche_store_pdp --version-id <version_id>

Key flags for `generate`:

• --agent-name — name for the agent (required)
• --prompt — natural language description of what to extract (required)
• --url — sample URL to analyze (required)
• --from-agent — existing agent to clone and refine (for iteration)
• --input-schema — custom input schema (optional, inferred if omitted)
• --output-schema — custom output schema (optional, inferred if omitted)
• --metadata — additional metadata (optional)

Generation response: returns id (generation ID), status (queued → in_progress → success / failed), and generated_version_id on success.

Workflow: Generate → poll with get-generation until success → optionally iterate with --from-agent → publish with version-id.

Polling: Generation takes 1-3 minutes. Run the generate → poll → publish loop as a background Task agent so the user isn't blocked waiting. The Task agent should poll nimble agent get-generation every 10 seconds until status is success or failed, then publish automatically (or report failure). Present results to the user when done.

MCP Fallback (when CLI is not installed)

If nimble --version returns "command not found", fall back to the Nimble MCP server. All CLI commands have MCP equivalents — discover them via the MCP tool list. MCP tools accept the same parameters as CLI flags, passed as tool arguments instead of flags.

Parallel Execution

Make multiple Bash tool calls in a single response. Claude Code runs them in parallel automatically:

• Call 1: nimble search --query "CompanyA news" --max-results 5 --search-depth lite
• Call 2: nimble search --query "CompanyB news" --max-results 5 --search-depth lite
• Call 3: nimble search --query "CompanyC news" --max-results 5 --search-depth lite

Sub-Agent Spawning

When using the Agent tool for parallel research:

• Always `mode: "bypassPermissions"` — sub-agents don't inherit Bash permissions.
• Batch max 4 agents. More risk hitting rate limits. For 5+, batch in groups.
• Tell agents to use Bash — explicitly say "Use the Bash tool to execute nimble

commands." Some agents try WebSearch instead.

• Fallback on failure — if any agent returns without results, run those searches

directly from the main context. Don't leave gaps.

Communication Style

Inform the user at phase transitions only with concrete numbers:

• "Researching Acme Corp + 5 competitors since Mar 12..."
• "Found 12 new signals. Pulling top 4 articles..."
• "All data collected. Building your briefing..."

Don't narrate individual tool calls.

Rate Limits & Common Errors

• Rate limit: 10 req/sec per API key
• Retry on 429: Reduce simultaneous calls
• Timeout: 30 seconds per request

Signal Date Validation

High-quality intelligence requires distinguishing between when a page was published and when the underlying event occurred. This matters because:

• Syndicated or republished content may carry a different publication date than the

original source

• Secondary coverage (regulatory filings, recap articles, industry roundups) can

report on events that happened weeks or months earlier

Article Date vs Event Date

Every signal has two dates:

A signal is "new" only if its event date falls within the freshness window.

Event Date Extraction Rules

Sub-agents must determine the event date from content:

1. Explicit past reference — "launched in Q3", "appointed last October" → event date is in the past, regardless of the article date 2. Temporal language — "last quarter", "months ago", "earlier this year" → resolve relative to the article date 3. Present tense announcement — "today announces", "is launching" → event date ≈ article date 4. Dateline — "NEW YORK, March 15 —" → event date = that dateline date 5. If ambiguous — extract the source URL and check the on-page date

Source Type Hierarchy

When the same event appears from multiple sources, prefer those closest to the event:

1. Primary — the company's own domain, official press release, regulatory filing 2. Wire service — AP, Reuters, Bloomberg 3. Major outlet — original reporting with bylines 4. Derivative — syndicated copies, aggregator sites, recap articles, or content that attributes its information to another source

If the only source for a signal is derivative, corroborate against a primary or major source before reporting.

Freshness Classification

After determining the event date, classify each signal:

Hard rule: Only signals classified as NEW or UPDATED may appear in reports. STALE and UNCERTAIN signals must be dropped entirely — not downgraded, not footnoted, not included as "background context." If a signal can't be verified as genuinely recent, it doesn't exist as far as the report is concerned.

--start-date Best Practices

--start-date is a useful filter for reducing noise, but always validate event dates from the content itself:

• For news queries (--focus news), consider running a parallel undated query to

surface original sources alongside recent coverage

• The existing fallback ("If < 3 results, retry without --start-date") remains useful

Verification Budget

Not every signal needs full verification — budget extract calls by priority:

Skills define their own P1/P2/P3 signal types in their SKILL.md. The verification budget above applies universally regardless of which signals a skill classifies at each level.

P1 Corroboration (Mandatory)

Any P1 signal sourced from derivative or aggregator sites must be corroborated before it can appear in a report. This is a hard gate, not a suggestion.

For each P1 signal that needs corroboration:

nimble search --query "[Company] [event summary]" --max-results 5 --search-depth lite

Look for the primary source (company blog, press release, official filing, regulatory document). Check the primary source's date:

• Primary source dates the event within the freshness window → signal is NEW, include it
• Primary source dates the event outside the freshness window → reclassify as STALE, drop
• No primary source found → reclassify as UNCERTAIN, drop

Do not report P1 signals that fail corroboration. It's better to miss a real signal than to report a stale one as new — trust is the product.

---

Entity Deduplication

When a skill collects entity records from multiple sources (directories, search results, extracted pages), deduplicate before reporting. This is distinct from signal-level differential analysis (see memory-and-distribution.md) — entity dedup merges records for the same entity across sources within a single run.

Three-layer pattern (generic — each skill customizes the specifics):

1. Exact ID match — If the entity type has a canonical ID (place_id, NPI number, domain), match on that first. Exact match = same entity, merge fields. 2. Domain normalization — Strip www., trailing slashes, protocol. Compare root domains. www.acme.com/ and acme.com are the same entity. 3. Fuzzy name + location — Normalize names before comparing:

• Lowercase all characters
• Strip titles and honorifics (Dr., Mr., Ms., etc.)
• Strip credential suffixes (MD, DDS, Inc, LLC, Corp, etc.)
• Strip common noise words (The, and, of, &)
• Collapse whitespace and punctuation
• Compare normalized names with location context if available

This catches cross-source variations like "Dr. Jane Smith, MD" (Maps) vs "Jane Smith" (Yelp) vs "Smith Eye Care LLC" (BBB). Each source formats names differently — always normalize before comparing.

Track source_count per entity — entities confirmed by multiple sources are higher quality. Each skill defines which layers apply and any domain-specific matching rules in its reference files.

---

Entity Confidence Scoring

Rate each entity's data completeness so users know what to trust.

Generic formula — each skill defines its own target field list (N fields):

• High — All target fields found + confirmed by 2+ sources (source_count >= 2)
• Medium — >50% of target fields found
• Low — ≤50% of target fields found

Display the confidence level in output (e.g., ⬤⬤⬤ High, ⬤⬤○ Medium, ⬤○○ Low). Each skill defines its field list and may add criteria (e.g., requiring a verified phone number for High in a provider directory skill).

---

Input Parsing Pattern

Skills that accept batch input (lists of URLs, companies, locations) should detect the input type automatically:

Normalize all inputs to a uniform list of records before batch processing. Don't assume a specific format — detect and adapt.

---

Scaled Execution

When a skill needs to run multiple WSA or API calls, choose the execution tier based on the estimated number of requests. Each skill calculates its own estimate from input size and operations per record.

Individual calls (1–10)

Run up to 4 concurrent Bash calls per the Parallel Execution rules above.

Batch calls (11+)

For page extraction (11+ URLs):

nimble extract-batch \
  --shared-inputs 'format: markdown' \
  --input '{"url": "https://example.com/page-1"}' \
  --input '{"url": "https://example.com/page-2"}'

Add --shared-inputs 'render: true' if pages need JavaScript rendering.

For WSA/agent calls (11+ entities):

nimble agent run-batch \
  --shared-inputs 'agent: {agent_name}' \
  --input '{"params": {...}}' \
  --input '{"params": {...}}'

Both return a batch_id. Poll progress:

nimble batches progress --batch-id {batch_id}

Fetch results when complete:

nimble batches get --batch-id {batch_id}
nimble tasks results --task-id {task_id}

Batch API handles up to 1,000 requests per call with server-side orchestration. For >1,000 requests, split into multiple batch calls.

Sub-agents should also batch. When spawning sub-agents for parallel work, tell each agent to use extract-batch or agent run-batch for its assigned items rather than making individual calls. One batch call per agent is faster and more reliable than 5-6 sequential calls.

Large job confirmation (>1,000)

Before executing, show the estimate and ask the user to confirm:

Estimated API calls: ~2,400 (120 locations × 3 WSAs per location × ~7 enrichment)
This is a large job. Proceed? [Y/n]

Pattern: estimate → display → gate → execute

Why batch over individual calls

Individual nimble agent run calls each require a separate HTTP round-trip and Bash tool invocation. At scale (dozens+) this is slow, unreliable, and wasteful on a local machine. Batch APIs move orchestration server-side — one API call triggers all requests, and you poll for results. Always prefer batch when above the individual threshold.

---

Query Construction Tips

• Be specific: "Acme Corp product launch 2026" > "Acme Corp"
• Use `--include-domain '["domain"]'` for companies with generic names
• Fallback on empty: If < 3 results, retry without --start-date
• Combine focus modes: news + general in parallel for broader coverage
• Try variations: "CompanyName" → "Company Name" → domain

NPI Verification Patterns

How to look up, interpret, and match NPI registry data against claimed practitioner credentials. For credential regex and specialty keywords, see provider-extraction-patterns.md. For general CLI patterns, see nimble-playbook.md.

---

NPI Lookup Strategy

The NPI (National Provider Identifier) is the authoritative source for verifying that a healthcare practitioner is registered, active, and practicing with the credentials they claim. Every individual healthcare provider in the US has a unique 10-digit NPI number.

Primary approach: NPPES API (1 call per provider)

The NPPES has a public API that returns structured JSON directly. This is faster and more reliable than search + extract (2 calls). Build the query URL:

nimble extract --url "https://npiregistry.cms.hhs.gov/api/?version=2.1&first_name=[First]&last_name=[Last]&state=[ST]&limit=5" --format markdown

Add &taxonomy_description=[Specialty] if the specialty is known. The response contains results[] with NPI number, status, credentials, taxonomy codes, addresses, and enumeration dates.

Parse the JSON response: The API returns escaped JSON in the markdown field. Parse it to extract the fields below. Match the correct result by comparing basic.first_name, basic.last_name, and taxonomies[].desc against claimed data.

Fallback approach: web search + extract (2 calls per provider)

If the NPPES API returns zero results or errors, fall back to web search:

nimble search --query "[Name] [Credential] [State] NPI registry" --max-results 5 --search-depth lite

Then extract from the top NPI source result:

nimble extract --url "[npi-result-url]" --format markdown

Source priority: 1. npidb.org/doctors/ — clean structured data 2. nppes.cms.hhs.gov (provider-view pages) — official CMS source 3. Skip all others (healthline, hmedata, vitals, etc.) — inconsistent formatting

Search refinement (if first search returns too many or zero results):

• Too many results (common names) — add specialty:

"[Name] [Specialty] [City] [State] NPI"

• No results — try without credentials:

"[Name] [State] NPI provider"

• Still no results — try with practice name:

"[Name] [Practice Name] NPI"

Search budget

Max 3 search queries + 1 extraction per practitioner. If no NPI match after 3 attempts, mark as Unverified and move on. The user can provide more context later for unresolved providers. This prevents runaway agents from burning 10+ calls on a single ambiguous name.

Key fields from NPI records

• NPI Number — 10-digit identifier
• Entity Type — Type 1 (individual) vs Type 2 (organization)
• Provider Name — legal name on record
• Credential — credential designation (MD, DO, OD, etc.)
• Taxonomy Code + Description — maps to specialty (see Taxonomy Matching below)
• Enumeration Date — when the NPI was first assigned
• Last Updated — most recent update to the record
• Status — Active or Deactivated (with deactivation date if applicable)
• Practice Address — primary practice location on file
• Mailing Address — may differ from practice address

---

Taxonomy Matching

NPI records use Healthcare Provider Taxonomy Codes to classify specialties. These are standardized codes maintained by the National Uniform Claim Committee (NUCC). Rather than maintaining a static lookup table, use the taxonomy description from the NPI record to match against the user's claimed specialty.

Matching strategy

1. Extract the taxonomy description from the NPI record (e.g., "Ophthalmology", "Optometrist", "Dentist - General Practice") 2. Normalize both the claimed specialty and NPI taxonomy description: lowercase, strip punctuation, expand abbreviations 3. Match levels:

• Exact — normalized strings match or one contains the other
• Partial — same broad category (e.g., claimed "retina specialist" vs

NPI taxonomy "Ophthalmology" — same field, subspecialty detail differs)

• Mismatch — different categories entirely (e.g., claimed "ophthalmologist"

vs NPI taxonomy "Dentist")

Common taxonomy categories by vertical

Use these to validate broad category matches when exact descriptions differ:

Ophthalmology family: Ophthalmology, Retina Specialist, Cornea Specialist, Glaucoma Specialist, Oculoplastics, Neuro-Ophthalmology, Pediatric Ophthalmology, Optometrist

Dental family: General Dentistry, Orthodontics, Periodontics, Endodontics, Oral Surgery, Prosthodontics, Pediatric Dentistry

Primary Care family: Family Medicine, Internal Medicine, General Practice, Geriatric Medicine, Preventive Medicine

The test for taxonomy matching: A "Partial" match within the same family is acceptable — flag it as a note, not an error. A mismatch across families is a verification failure that requires human review.

---

Verification Status Determination

For each practitioner, assign one of four verification statuses based on the totality of evidence:

Verified

All of the following are true:

• NPI record found and status is Active
• Name matches (after normalization — see name matching below)
• Credential matches (claimed credential appears on NPI record)
• Specialty matches (Exact or Partial taxonomy match)
• Practice address is in the claimed state (if state was provided)

Partially Verified

NPI record found but with minor discrepancies:

• Name matches but credential not listed on NPI (may use different designation)
• Specialty is a Partial match (same family, different subspecialty)
• Address is in a different city but same state
• NPI is active but hasn't been updated in 5+ years

Unverified

Cannot confirm the practitioner's claims:

• No NPI record found for this name + state combination
• Multiple NPI matches, insufficient data to disambiguate
• NPI record found but name similarity is too low (possible different person)

Flagged

Active discrepancies that require human review:

• NPI record exists but status is Deactivated (include deactivation date)
• Credential mismatch — claimed MD but NPI shows DO, or vice versa
• Specialty mismatch across families (claimed ophthalmologist, NPI says dentist)
• Address mismatch — claimed state differs from NPI practice state
• Name match is borderline (could be same person with name change, or different person)

Recent Relocators

If the NPI address is in a different state but the provider's practice website shows a recent join date (< 6 months), classify as Verified with note rather than Partially Verified or Flagged. NPI address updates typically lag 3-6 months after a provider changes practice locations. Note the discrepancy but don't downgrade the verification status. Look for these signals:

• Practice website bio says "joined [practice] in [recent year]"
• Provider appears on the new practice's team page but NPI shows prior state
• LinkedIn or other sources confirm the relocation

---

Name Matching

NPI records use legal names which may differ from how providers present themselves on practice websites.

Normalization steps (apply to both claimed name and NPI name)

1. Strip titles: Dr., Mr., Ms., Prof. 2. Strip credentials: all patterns from provider-extraction-patterns.md 3. Strip suffixes: Jr., Sr., III, IV 4. Lowercase everything 5. Collapse whitespace, strip remaining punctuation

Match levels

• Strong match — normalized first + last name match exactly
• Likely match — last name matches, first name is a known variant

(e.g., "William" / "Bill", "Robert" / "Bob", "James" / "Jim")

• Weak match — last name matches, first name differs (possible name change,

or different person — flag for review)

• No match — last names differ after normalization

Common name variant pairs

Rather than hardcoding a complete list, look for these patterns:

• Formal vs informal (William/Bill, Robert/Bob, James/Jim, Elizabeth/Beth)
• Hyphenated vs maiden name (provider may use married or maiden name differently

across sources)

• Middle name as first name (some providers go by their middle name)

When a name is a "Weak match" or "No match" but other fields strongly match (same specialty, same city, same practice name), upgrade to "Likely match" and note the discrepancy rather than marking as Unverified.

---

Mismatch Flagging

Every flagged mismatch must include: 1. What was claimed — the value from the user's input 2. What was found — the value from the NPI record 3. Source URL — clickable link to the NPI lookup page 4. Severity — Critical (deactivated NPI, credential mismatch) or Warning (address mismatch, stale record, name variant)

Critical mismatches (always flag)

• Deactivated NPI status
• Credential type mismatch (MD claimed, DO found — or vice versa)
• Specialty mismatch across taxonomy families
• Entity type mismatch (individual claimed, organization NPI found)

Warning mismatches (flag as notes)

• Address in different city but same state
• Subspecialty mismatch within same taxonomy family
• NPI record not updated in 5+ years (stale but not necessarily wrong)
• Name variant (informal vs formal name)

---

Batch Verification Optimization

For large lists (20+ practitioners), optimize the search strategy:

Group by state

Practitioners in the same state often appear in the same NPI lookup results. Group searches by state to improve cache hits and reduce redundant searches.

Prioritize by risk

If the user specified a verification focus (e.g., "check credentials only"), optimize the pipeline:

• Credentials only — NPI lookup + credential match, skip address verification
• Active status only — NPI lookup + status check, skip taxonomy matching
• Full verification — all checks (default)

Checkpoint strategy

Save verification progress after every 10 practitioners:

echo '{...}' > ~/.nimble/memory/healthcare-providers-verify/checkpoints/{slug}/verify.json

This allows resuming interrupted verification runs without re-searching already-verified practitioners.

Profile & Onboarding

The business profile at ~/.nimble/business-profile.json and first-run setup flow.

---

Profile Schema

{
  "company": {
    "name": "Acme Corp",
    "domain": "acme.com",
    "description": "Enterprise SaaS platform for project management"
  },
  "industry_keywords": ["project management software", "team collaboration SaaS"],
  "competitors": [
    { "name": "WidgetCo", "domain": "widgetco.com", "category": "project-mgmt" },
    { "name": "GizmoTech", "domain": "gizmotech.io", "category": "project-mgmt" }
  ],
  "preferences": {
    "skip_competitors": [],
    "output_format": "bullet-points"
  },
  "integrations": {
    "notion": { "reports_page_id": "" },
    "slack": { "channel": "" }
  },
  "sales_context": {
    "key_differentiators": [
      "Only platform with real-time web data access",
      "Sub-second API response times"
    ],
    "integration_partners": [
      { "name": "DataStack", "type": "data warehouse" },
      { "name": "CRMHub", "type": "CRM" }
    ],
    "case_studies": [
      { "customer": "Large enterprise retailer", "industry": "retail", "outcome": "3x faster competitive intel" }
    ],
    "common_objections": [
      { "objection": "We already use [competitor]", "response": "Our real-time data is fresher — most competitors cache for 24h+" }
    ]
  },
  "last_runs": {
    "competitor-intel": "2026-03-20T14:30:00Z",
    "meeting-prep": "2026-03-22T09:00:00Z"
  },
  "setup_completed": true
}

Reading the Profile

At the start of every skill run:

cat ~/.nimble/business-profile.json 2>/dev/null

If missing or empty → trigger onboarding (see below).

Key fields:

• company.name / company.domain — the user's company
• competitors — tracked competitors with domains and categories
• industry_keywords — for industry-level searches
• preferences.skip_competitors — competitors to exclude
• last_runs.{skill-name} — timestamp for time-aware searches
• sales_context — value positioning data (differentiators, integrations, case studies, objections)
• integrations — Notion/Slack config for report distribution

Updating the Profile

After every skill run — update last_runs:

import json, datetime, os
path = os.path.expanduser("~/.nimble/business-profile.json")
with open(path, "r") as f:
    profile = json.load(f)
profile["last_runs"]["skill-name"] = datetime.datetime.now(datetime.timezone.utc).isoformat()
with open(path, "w") as f:
    json.dump(profile, f, indent=2)

On user correction — apply immediately:

Always confirm: "Got it — removed CompanyX from tracking."

Rules:

• Never overwrite the whole file. Read → modify → write.
• Preserve unknown fields.
• Handle missing file gracefully → trigger onboarding.
• JSON only, always valid.

---

First-Run Onboarding

Prerequisite Checks

The transport selection in nimble-playbook.md determines whether CLI or MCP is active. This section covers the install/upgrade/auth flow when neither is ready.

Minimum CLI version: 0.12.0

Preferred path — any Claude product (Claude Code, Claude Cowork, claude.ai)

The plugin install is one command and handles MCP registration + OAuth automatically:

"Run /plugin install nimble to install the Nimble plugin. The plugin's MCP

server auto-registers as a Connector you can see in Customize → Connectors.

On first use, the OAuth flow runs in your browser — no API key needed."

This works in every Claude product (Code, Cowork, claude.ai) — they share the plugin + connector mechanism.

Plugin installed but connector not connected (Cowork / claude.ai)

The most common Cowork / claude.ai failure: the plugin is installed (mcp__plugin_nimble_nimble__* tools are listed) but its connector isn't connected, so live data calls fail. Check this before doing any work — don't fire a data call and react to the error. A single read-only nimble_agents_list probe confirms it: success = connected, proceed; auth/not-connected error or a response containing an OAuth authorization URL = not connected.

When not connected, tell the user verbatim and stop — never fall back to WebFetch, WebSearch, or any other tool, and never guess at data:

Your Nimble plugin is installed, but its connector isn't connected yet — that's

why live data isn't working. To connect it:

>

1. Open Customize → Connectors

2. Find Nimble and click Connect

3. Complete the login in your browser. No Nimble account? You can create one

right there during login.

4. Once it shows Connected, re-run your request.

If a tool returns an OAuth "Authorize" link instead of data, present the link as-is and stop. Do not invent a completion step ("paste the URL back", "I'll complete the connection") — no such step exists. Do not claim the tools will activate and then call them in the same turn. Wait for the user to authorize, then retry (or run one nimble_agents_list probe to confirm).

Codex CLI or other terminal agents (shell available, no /plugin install)

When /plugin install isn't available but the user has shell access, install the CLI directly — it exposes the full Nimble surface area:

1. Check if npm is available: npm --version 2. If npm exists:

"The Nimble CLI is required. I'll install it now."

>

Run: npm install -g @nimble-way/nimble-cli

3. If npm is not available:

"The Nimble CLI requires Node.js/npm. Install Node.js first from

nodejs.org, then run: npm install -g @nimble-way/nimble-cli"

4. After install, verify: nimble --version 5. If verification fails, stop and ask the user to check their PATH.

Cursor, VS Code, or other MCP clients outside the Claude family

When neither /plugin install nor shell access is workable, have the user paste this into their MCP settings (e.g., .cursor/mcp.json or the host's equivalent):

{
  "mcpServers": {
    "nimble": {
      "type": "http",
      "url": "https://mcp.nimbleway.com/mcp"
    }
  }
}

After install, the first tool call triggers the OAuth flow automatically.

CLI outdated (version < 0.12.0)

Parse the version from nimble --version. If below 0.12.0:

"Your Nimble CLI is version [current] — version 0.12.0+ is required

for these skills. Upgrading now..."

>

Run: npm update -g @nimble-way/nimble-cli

Verify after upgrade: nimble --version. If still outdated, suggest: npm uninstall -g @nimble-way/nimble-cli && npm install -g @nimble-way/nimble-cli

API key not set

You need a Nimble API key.

1. Go to app.nimbleway.com → API Keys

2. Generate a new key

3. Run: export NIMBLE_API_KEY=your_key_here

4. Add to ~/.zshrc or ~/.bashrc to make permanent.

After the user sets it, verify: echo "NIMBLE_API_KEY=${NIMBLE_API_KEY:+set}"

API key expired (401)

Your key may have expired (72h TTL). Regenerate at app.nimbleway.com > API Keys.

All prerequisites met

Only proceed to Company Setup once CLI is installed, version is >= 0.12.0, and API key is set. Don't silently skip any check.

Company Setup (2 prompts max)

Prompt 1 — ask in plain text (NOT AskUserQuestion with options):

"What's your company's website domain? (e.g., acme.com)"

Verify — make two Bash calls simultaneously:

• nimble search --query "[domain]" --include-domain '["[domain]"]' --max-results 3 --search-depth lite
• nimble search --query "[domain] company" --max-results 5 --search-depth lite

Present what you found and confirm: "I found that [Company] ([domain]) is [brief description]. Is this the right company?"

Prompt 2 — skill-specific setup:

• competitor-intel: Offer choice via AskUserQuestion:
• Find for me — search and suggest competitors
• I'll list them — user provides names

If "Find for me", make three Bash calls simultaneously:

• nimble search --query "[Company] competitors" --max-results 10 --search-depth lite
• nimble search --query "[Company] vs" --max-results 10 --search-depth lite
• nimble search --query "[Company] alternatives" --max-results 5 --search-depth lite

• meeting-prep: No extra setup — context comes per-meeting
• company-deep-dive: No extra setup — target company comes per-request

Create Profile

mkdir -p ~/.nimble/memory/{competitors,people,companies,reports,positioning,synthesis}

Write ~/.nimble/business-profile.json using the schema above.

When setting up competitors, infer or ask for each competitor's domain and category. Also infer industry keywords from the company description.

Profile Exists

Skip onboarding. Greet with context: "Running competitor intel for Acme Corp — tracking WidgetCo, GizmoTech."

---

Error Recovery

If any step fails: 1. Tell the user what went wrong in plain language 2. Provide the exact command to fix it 3. Offer to retry

Never silently skip setup steps.

Provider Extraction Patterns

Skill-specific patterns for extracting practitioner data from healthcare practice websites. For general extraction and site mapping rules, see nimble-playbook.md.

---

Page URL Scoring

After nimble map discovers a site's pages, score each URL by path keywords to identify provider-relevant pages. See Site Mapping Pattern in nimble-playbook.md for the generic discover/score/filter/fallback flow.

Keyword Weight Table

Scoring Rules

• Cap at 15 pages per practice site
• Always include at least one High-weight page (or homepage as fallback)
• Individual provider pages (/dr-*) are high value but cap at 10 per site to

avoid over-extraction on large multi-provider practices

• If nimble map returns < 3 scored candidates, fall back to:

  nimble search --query "site:{domain} doctors OR providers OR team" --max-results 10 --search-depth lite

---

Core Extraction Fields (5 Fields)

Every provider record targets these 5 fields. Confidence scoring uses this as the N-field list (see Entity Confidence Scoring in nimble-playbook.md).

Confidence Scoring (from shared pattern)

• High -- 5/5 fields found + confirmed by 2+ pages or sources
• Medium -- 3-4/5 fields found
• Low -- 1-2/5 fields found

Display as: High, Medium, Low

---

Credential Regex Patterns

Match these suffixes adjacent to provider names. Case-insensitive. Allow comma, space, or period separators between multiple credentials.

Medical Doctors

MD|M\.D\.|DO|D\.O\.

Eye Care

OD|O\.D\.|FAAO

Dental

DDS|D\.D\.S\.|DMD|D\.M\.D\.

Advanced Practice

NP|PA|PA-C|ARNP|APRN|CNS|CRNA|DNP|D\.N\.P\.

Therapy & Allied Health

PT|DPT|OT|OTR|SLP|CCC-SLP|RD|RDN|LCSW|LPC|PhD|Ph\.D\.|PsyD|Psy\.D\.

Board Certifications (commonly listed)

FACS|FACP|FACC|FACOG|FAAP|FACEP|FAAOS|FASRS|FRCSC

Combined Pattern

When scanning extracted markdown, look for names followed by credential clusters:

[Name],?\s*((?:MD|DO|OD|DDS|DMD|NP|PA|PA-C|ARNP|APRN|PT|DPT|PhD|FACS|FAAO|FACP|FACC|FACOG|FAAP|FAAOS|FASRS)[,.\s]*)+

---

Specialty Keywords by Healthcare Vertical

Ophthalmology

ophthalmology, ophthalmologist, retina, retinal, cataract, glaucoma, cornea,
corneal, LASIK, refractive surgery, oculoplastics, neuro-ophthalmology,
pediatric ophthalmology, vitreoretinal, anterior segment, posterior segment,
strabismus, ocular oncology, uveitis

Dental

dentist, dentistry, general dentistry, cosmetic dentistry, orthodontics,
orthodontist, periodontics, periodontist, endodontics, endodontist,
oral surgery, oral surgeon, prosthodontics, prosthodontist, pediatric
dentistry, implants, dental implants, TMJ, sedation dentistry

Dermatology

dermatology, dermatologist, Mohs surgery, cosmetic dermatology, skin cancer,
melanoma, psoriasis, eczema, acne, rosacea, laser treatment, botox,
fillers, chemical peel, phototherapy, patch testing

General / Primary Care

family medicine, internal medicine, primary care, general practice,
preventive medicine, geriatrics, urgent care, walk-in clinic,
physical exam, wellness, annual checkup

Orthopedics

orthopedics, orthopedic surgery, sports medicine, joint replacement,
spine surgery, hand surgery, foot and ankle, shoulder, knee,
arthroscopy, fracture care, physical therapy, rehabilitation

---

Entity Deduplication (Skill-Specific)

Apply the shared 3-layer dedup pattern from nimble-playbook.md with these skill-specific rules:

1. Exact match -- Same name + same practice domain = same provider 2. Credential match -- Same name + same credentials + same city = likely same provider (even across different practice sites) 3. Fuzzy match -- Normalize names (strip "Dr.", middle initials, suffixes), compare with Levenshtein distance <= 2 + same specialty = possible match, flag for review rather than auto-merging

Cross-source name normalization

Different sources format provider names very differently. Exact string matching across sources produces near-zero matches. Always normalize before comparing:

Normalization steps (apply in order): 1. Strip titles: Dr., Mr., Ms., Prof. 2. Strip credentials: all patterns from the Credential Regex section above 3. Strip business suffixes: LLC, Inc, Corp, PC, PLLC, PA, Associates 4. Strip specialty descriptors: "- Retina Specialist", "- Ophthalmologist" 5. Strip middle initials (single letters with optional period) 6. Lowercase, collapse whitespace, strip remaining punctuation

After normalization, match with location context (same city or same zip code). For practice-level dedup (not provider-level), also try matching the practice name against provider last names ("Smith Eye Care" → likely matches "Dr. Smith").

Track source_count -- providers found across multiple sources are higher confidence than those from a single source.

WSA Discovery for Healthcare Providers Verify

How to find and evaluate WSAs for validating practitioner credentials and license status. The WSA catalog evolves constantly — this skill discovers relevant agents at runtime rather than relying on a static list.

For general WSA execution rules (invocation, parsing, batch, fallback), see nimble-playbook.md.

---

Discovery Strategy

Three search layers

Run these searches during preflight (Step 0) to build a session-specific WSA inventory. Run all searches simultaneously:

Layer 1 — Vertical search:

nimble agent list --limit 100 --search "healthcare"

Returns all agents tagged with the Healthcare vertical (clinical trials, FDA, regulatory, licensing, and any newly added healthcare agents).

Layer 2 — Session-specific search: Search for terms derived from the user's input — their specialty, specific registries, or data sources they mentioned:

# Registry and licensing searches:
nimble agent list --limit 50 --search "npi"
nimble agent list --limit 50 --search "license"
nimble agent list --limit 50 --search "board certification"

# If user mentioned specific registries:
nimble agent list --limit 50 --search "medical board"
nimble agent list --limit 50 --search "[state] license lookup"

# Specialty-specific:
nimble agent list --limit 50 --search "[user's specialty]"

Adapt search terms to whatever the user provided. Focus on regulatory and verification-related agents.

Layer 3 — General verification tools: These WSAs are useful across verticals for identity confirmation and practice verification:

nimble agent list --limit 50 --search "google_maps"
nimble agent list --limit 50 --search "yelp"
nimble agent list --limit 50 --search "bbb"

Evaluating discovered agents

For each discovered agent, read its description and entity_type to classify it into a verification category:

Validate each relevant agent's params before using it:

nimble agent get --template-name [agent_name]

Skip agents that don't fit — not every healthcare-tagged agent is useful for verification. A drug interaction agent or a clinical trial search agent adds no value to credential validation.

---

Verification Phase Mapping

This skill has a narrower WSA focus than extract or enrich — the primary tool is nimble search + nimble extract for NPI lookups. WSAs add value in two areas:

Credential verification

When: Always — this is the core value of the skill.

Primary tool: nimble search + nimble extract (not WSA-dependent):

nimble search --query "[name] [credential] [state] NPI registry" --max-results 5 --search-depth lite

Then extract the NPI result page for structured data.

WSA supplement: If Layer 1/2 discovery found NPI or medical board agents, use them for direct lookups — structured WSA output is higher quality than parsing search results.

Regulatory verification

When: User wants to confirm clinical trial activity, FDA associations, or regulatory compliance beyond basic NPI lookup.

Search terms for discovery: healthcare vertical, clinicaltrials, fda, board, license

How to use: Run discovered regulatory agents with the provider's name and credentials. Cross-reference with NPI data to build a fuller verification picture.

Fallback:

nimble search --query "[provider-name] [credentials] clinical trials OR FDA OR board certification" --max-results 5 --search-depth lite

Practice confirmation

When: Need to confirm a provider is actively practicing at a claimed location.

Search terms for discovery: google_maps, yelp, bbb

How to use: Run practice-level agents with the provider's practice name + location. Confirm the practice exists and is active.

Fallback:

nimble search --query "[practice-name] [city] [state]" --max-results 5 --search-depth lite

---

Scaling WSA Calls

Follow the Scaled Execution pattern from nimble-playbook.md — it covers individual calls, batching, and the confirmation gate for large jobs.

For verification, the primary bottleneck is NPI lookups (one per practitioner). WSA calls add supplementary verification and scale as: [practitioners requesting regulatory/practice verification] x [relevant WSA categories].

---

Fallback Chain

If WSA discovery returns nothing useful for a category, fall back to nimble search + nimble extract (the core Nimble tools always work):

1. Credential fallback: nimble search --query "[name] [state] NPI" --max-results 5 --search-depth lite + extract NPI result 2. Regulatory fallback: nimble search --query "[name] [credentials] clinical trials OR FDA" --max-results 5 --search-depth lite 3. Practice fallback: nimble search --query "[practice-name] [city] [state] reviews" --max-results 5 --search-depth lite

The skill always produces verification results even if zero WSAs are found — WSAs accelerate and enrich the verification, but NPI lookup via web search is the foundation and never requires WSAs.