Doctorg

Doctor G -- Evidence-Based Health Research

Answer health and wellness questions using only trusted, evidence-based sources with explicit evidence strength ratings.

Usage

# Quick answer (WebSearch only, ~30s)
/doctorg Is creatine safe for daily use?

# Deep research (WebSearch + Tavily, ~90s)
/doctorg --deep Huberman vs Attia on fasted training

# Full investigation (WebSearch + Tavily + Firecrawl, ~3min)
/doctorg --full What does current evidence say about GLP-1 agonists for non-diabetic weight loss?

# Without personal health context
/doctorg --no-personal Best stretching protocol for lower back pain

Depth Levels

How It Works

1. Parse Query & Detect Topic Category

Classify the question into one of:

• Nutrition/Supplements (examine.com gets priority)
• Exercise/Training (PubMed + ACSM get priority)
• Sleep (focus sleep-specific databases)
• Disease/Condition (condition-specific orgs + clinical guidelines)
• Medication/Treatment (FDA, EMA, Cochrane get priority)
• Mental Health (APA, mental health orgs)
• General Wellness (broad search across all tiers)

2. Search Evidence Sources (Tiered)

Search sources in priority order. See references/sources.md for complete domain list.

Tier 1 -- Primary Research (highest weight):

• PubMed/PMC, Cochrane Library, WHO, ClinicalTrials.gov

Tier 2 -- Clinical/Institutional (high weight):

• Mayo Clinic, Hopkins Medicine, Cleveland Clinic, Harvard Health
• Condition-specific: AHA, ACS, ADA, Alzheimer's Association

Tier 3 -- Expert Analysis (medium weight):

• Examine.com, STAT News, Health News Review
• Consensus.app, Epistemonikos

Tier 4 -- Quality Journalism (context/framing):

• The Atlantic, NYT, NPR, Guardian, FiveThirtyEight

Search Strategy by Depth

Quick (default):

WebSearch(query, allowed_domains=[Tier 1 + Tier 2 domains])
WebSearch(query + "systematic review OR meta-analysis", allowed_domains=[Tier 1])

Deep (--deep): All Quick searches PLUS:

tavily-search(query, include_domains=[Tier 1-3])
WebSearch(query + "expert opinion OR position statement", allowed_domains=[Tier 2-3])
WebSearch(query + "risks OR side effects OR contraindications")

Full (--full): All Deep searches PLUS:

firecrawl-research for top 2-3 most relevant results from Tier 1
WebSearch for competing/contrarian viewpoints
WebSearch(query + "retracted OR debunked OR misleading")

3. Pull Personal Health Context (unless --no-personal)

Query Apple Health database for relevant metrics:

python ~/ai_projects/claude-skills/health-data/scripts/health_query.py --format json vitals
python ~/ai_projects/claude-skills/health-data/scripts/health_query.py --format json daily
python ~/ai_projects/claude-skills/health-data/scripts/health_query.py --format json sleep --days 7
python ~/ai_projects/claude-skills/health-data/scripts/health_query.py --format json workouts --days 30

Select ONLY metrics relevant to the query:

• Exercise question -> recent workouts, activity, resting HR, VO2 max
• Sleep question -> sleep data, HRV
• Nutrition question -> weight trends, activity level
• Heart question -> HR, HRV, resting HR, blood pressure

4. Synthesize with Evidence Grading

Rate each claim using simplified GRADE scale:

5. Format Output

# [Topic Title]

**Short answer**: [1-2 sentence direct answer]

## [Expert/Position A] (if comparing viewpoints)
- Key claim 1
- Key claim 2
- Has **evolved stance**: [if applicable]

## [Expert/Position B]
- Key claim 1
- Key claim 2

## Where They Actually Agree (if comparing)
- Agreement point 1
- Agreement point 2

## What Research Shows

| Claim | Evidence Strength |
|-------|------------------|
| Claim 1 | **Strong** |
| Claim 2 | **Weak** (reason) |
| Claim 3 | **Contested** |

## For You Specifically (if --personal context available)

[Personalized interpretation based on user's health data]

[Specific actionable recommendation]

## Sources
- [Source 1 title](url) -- Tier, year
- [Source 2 title](url) -- Tier, year

## Limitations
- [Any caveats about the evidence or this analysis]

Output rules:

• NEVER give medical diagnoses or replace professional advice
• ALWAYS include disclaimer: "This is research synthesis, not medical advice"
• When evidence is Weak or Minimal, explicitly say so
• When claims are Contested, present both sides fairly
• Prefer recent sources (last 5 years) over older ones
• Flag if key studies have been retracted or challenged
• Include the "For You Specifically" section only when health data adds meaningful context

6. Disclaimer (always append)

---
*Research synthesis, not medical advice. Consult a healthcare provider for personal decisions.*

Examples

Quick

/doctorg Is 10000 steps a day backed by science?

Deep (comparing experts)

/doctorg --deep Huberman vs Attia on fasted training

Full (controversial topic)

/doctorg --full Safety profile of long-term melatonin supplementation

Integration with Other Skills

• health-data: Pulls Apple Health metrics for personalization
• tavily-search: Deep research at Tier 1-3 sources
• firecrawl-research: Full-text extraction from primary sources
• fact-checker: Can be chained for verification of specific claims

Doctor G -- Implementation Instructions

Invocation

/doctorg [--deep|--full] [--no-personal] <question>

Parse the argument to extract:

• depth: "quick" (default), "deep" (--deep), or "full" (--full)
• personal: true (default), false (--no-personal)
• question: everything after flags

Step-by-Step Execution

Step 1: Classify the Question

Determine question type to select source priorities:

Pattern matching:

• Contains supplement/vitamin/mineral names -> nutrition_supplements
• Contains exercise/training/workout/lifting -> exercise_training
• Contains sleep/insomnia/circadian -> sleep
• Contains heart/cardio/blood pressure -> cardiovascular
• Contains cancer/tumor/oncology -> oncology
• Contains diabetes/insulin/glucose/blood sugar -> diabetes
• Contains anxiety/depression/mental/therapy -> mental_health
• Contains drug/medication/prescription -> medication
• Contains weight/fat loss/body comp/BMI -> body_composition
• Contains "vs" or "versus" or "compared to" -> expert_comparison (add flag)
• Default -> general_wellness

Expert comparison detection: If question contains "vs", "versus", or names two known health figures (Huberman, Attia, Rhonda Patrick, Greger, etc.), set is_comparison = true.

Step 2: Build Search Queries

For each depth level, construct search queries.

All levels -- base queries:

queries = [
    f"{question}",
    f"{question} systematic review OR meta-analysis",
]

Deep -- additional queries:

queries += [
    f"{question} risks OR side effects OR limitations",
    f"{question} expert consensus OR position statement",
    f"{question} evidence strength",
]

Full -- additional queries:

queries += [
    f"{question} retracted OR debunked OR criticism",
    f"{question} recent 2025 2026 update",
    f"{question} contrarian view OR counterargument",
]

Step 3: Execute Searches

Quick depth: 1. WebSearch(query=base_query, allowed_domains=[Tier1 + Tier2 + topic-specific Tier3]) 2. WebSearch(query=systematic_review_query, allowed_domains=[Tier1])

Deep depth (run searches in parallel where possible): 1. All Quick searches 2. WebSearch(query=risks_query, allowed_domains=[Tier1 + Tier2]) 3. WebSearch(query=expert_query, allowed_domains=[Tier2 + Tier3]) 4. Use tavily-search skill: tavily-search "{question}" --include-domains {Tier1+Tier2+Tier3 comma-separated} 5. If is_comparison: search each expert's name + topic separately

Full depth: 1. All Deep searches 2. WebSearch(query=contrarian_query) (no domain filter -- find opposing views) 3. WebSearch(query=retraction_query, allowed_domains=[Tier1]) 4. Use firecrawl-research to extract full text from top 2-3 most relevant Tier 1 results 5. If comparison: firecrawl-research on each expert's primary source (podcast transcript, blog post, paper)

Step 4: Pull Personal Health Context

Skip if `--no-personal` flag set.

Run relevant health queries based on topic:

# Always pull (baseline context)
python ~/ai_projects/claude-skills/health-data/scripts/health_query.py --format json vitals

# Topic-specific
# exercise/training:
python ~/ai_projects/claude-skills/health-data/scripts/health_query.py --format json workouts --days 30
python ~/ai_projects/claude-skills/health-data/scripts/health_query.py --format json activity --days 7

# sleep:
python ~/ai_projects/claude-skills/health-data/scripts/health_query.py --format json sleep --days 14

# body composition/nutrition:
python ~/ai_projects/claude-skills/health-data/scripts/health_query.py --format json weekly --weeks 4

# cardiovascular:
python ~/ai_projects/claude-skills/health-data/scripts/health_query.py --format json vitals
# + custom query for BP if available:
python ~/ai_projects/claude-skills/health-data/scripts/health_query.py --format json query "SELECT AVG(value), unit FROM health_records WHERE record_type LIKE '%BloodPressure%' AND start_date >= date('now', '-30 days')"

Health context formatting: Extract only the relevant numbers. Don't dump raw JSON into the response. Example: "Based on your recent data: resting HR 62 bpm, HRV 45ms, avg 9,200 steps/day, 4 strength workouts/week"

Step 5: Synthesize Evidence

Evidence grading rules: 1. Start with study design level (see sources.md grading table) 2. Upgrade if: large effect size, dose-response, multiple independent replications 3. Downgrade if: industry funding without replication, small sample, high heterogeneity, indirect evidence 4. Strong: 2+ systematic reviews/meta-analyses agree, or 3+ large RCTs consistent 5. Moderate: 1 systematic review or 2+ well-designed studies, some limitations 6. Weak: Limited studies, small samples, or conflicting results 7. Minimal: Expert opinion only, animal studies, case reports 8. Contested: Credible evidence on both sides, active scientific debate

For expert comparisons: 1. State each expert's position clearly with their key claims 2. Find where they actually agree (usually more than expected) 3. Show what research supports/contradicts each position 4. Avoid declaring a "winner" -- let evidence speak

Step 6: Format Output

Follow the template in SKILL.md exactly. Key formatting rules:

1. Short answer first -- 1-2 sentences, direct 2. Evidence table -- every major claim gets a strength rating 3. Personal section -- only if health data adds meaningful context 4. Sources -- list with tier label and year 5. Disclaimer -- always append

Table formatting:

| Claim | Evidence Strength |
|-------|------------------|
| Creatine improves strength output | **Strong** |
| Creatine causes hair loss | **Weak** (single small study, not replicated) |
| 5g/day is optimal dose | **Strong** |
| Loading phase is necessary | **Weak** (not needed, just takes longer to saturate) |

Personal section example:

## For You Specifically

Your recent data shows 4 strength sessions/week with avg resting HR of 62.
Given your training volume, 5g creatine monohydrate daily would be well-supported.
Your HRV of 45ms suggests good recovery capacity.

**One consideration**: Your sleep data shows avg 6.2h -- optimizing sleep to 7+ hours
would likely provide more performance benefit than any supplement.

Error Handling

Database not found (health.db):

• Skip personal context section
• Note: "Personal health data unavailable -- using general recommendations"

No relevant search results from Tier 1:

• Expand to Tier 2-3
• Note: "Limited primary research found -- evidence grading reflects available sources"

Question outside health domain:

• Politely redirect: "This skill focuses on evidence-based health questions. For [topic], try [alternative]."

Highly personalized medical question (e.g., "should I take this medication"):

• Provide evidence but emphasize: "This specific decision requires your healthcare provider who knows your full history"

Performance Notes

• Quick depth should return in <30s (2 WebSearch calls)
• Deep depth: ~60-90s (5-6 search calls + tavily)
• Full depth: ~2-3min (8+ searches + firecrawl)
• Health data queries: ~2-3s each (local SQLite)
• Parallelize independent search calls where possible

Evidence Source Registry

Tiered domains for WebSearch allowed_domains parameter and Tavily include_domains.

Tier 1 -- Primary Research & Systematic Reviews

Highest evidence weight. Peer-reviewed, systematic.

Databases & Aggregators

WebSearch Domain List (copy-paste for Tier 1)

ncbi.nlm.nih.gov, cochranelibrary.com, who.int, clinicaltrials.gov, europepmc.org, epistemonikos.org

Tier 2 -- Clinical & Institutional

High evidence weight. Expert institutions, clinical guidelines.

General Medical

Condition-Specific Organizations

Regulatory

WebSearch Domain List (copy-paste for Tier 2)

mayoclinic.org, hopkinsmedicine.org, clevelandclinic.org, health.harvard.edu, medlineplus.gov, nhlbi.nih.gov, cdc.gov, nih.gov, nice.org.uk, americanheart.org, cancer.org, cancer.gov, diabetes.org, alz.org, fda.gov, ema.europa.eu, nimh.nih.gov

Tier 3 -- Expert Analysis & Evidence Synthesis

Medium weight. Curated expert analysis, evidence-based summaries.

AI-Powered Research (use as supplementary)

WebSearch Domain List (copy-paste for Tier 3)

examine.com, statnews.com, healthnewsreview.org, healthfeedback.org, acsm.org, nsca.com, sleepfoundation.org, consensus.app

Tier 4 -- Quality Journalism

Lowest evidence weight. Use for context, framing, public discourse. Never cite as primary evidence.

WebSearch Domain List (copy-paste for Tier 4)

theatlantic.com, nytimes.com, npr.org, theguardian.com, fivethirtyeight.com

Topic-Specific Source Priority

When the question is about a specific topic, prioritize these sources:

Evidence Grading Quick Reference

Red Flags to Watch For

When evaluating sources, flag these:

• Industry-funded studies without independent replication
• Retracted papers (check Retraction Watch)
• Predatory journals (check DOAJ for journal quality)
• Supplements with proprietary blends (can't verify doses)
• N=1 or case report presented as generalizable
• Animal studies presented as human-applicable
• Relative risk without absolute risk context
• Correlation presented as causation