Postplus Shared

PostPlus Shared

Use this skill first when another PostPlus skill asks for shared rules or shared workflow preferences.

This skill is the single source of truth for principle-level PostPlus skill documents. It is instruction-only: do not run provider runtimes, mutate public metadata or release metadata, or invent alternate execution paths from this shared-rule handoff.

References

• `shared-public-skill-rules.md`
• `shared-research-preferences.md`
• `shared-product-selection-preferences.md`
• `shared-source-of-truth-files.md`
• `shared-tiktok-music-workflow.md`
• `shared-ads-workflow.md`
• `shared-user-guidance.md`

Default Workflow

1. Read the reference named by the downstream skill. 2. Apply the shared rule before reading platform- or production-specific details. 3. If a downstream skill and this shared rule conflict, fail fast and surface the conflict instead of inventing a fallback. 4. Keep executable work in the downstream skill or PostPlus CLI boundary; this shared skill only owns the principle-level rule.

Shared Ads Workflow

Shared routing and safety rules for the ad skill family.

Use this file when a task touches any ad platform and you need the common model before platform-specific work.

Core Rule

Do not treat ad management as a single free-form agent task.

Split it into five stages:

1. connect account 2. read data 3. normalize 4. diagnose and propose 5. approve and apply

This family should always keep those stages separate.

Default Flow

flowchart LR
    A["User intent"] --> B["ads-router"]
    B --> C["connect / discover"]
    C --> D["read / normalize"]
    D --> E["diagnose"]
    E --> F["propose patch"]
    F --> G["human approval gate"]
    G --> H["apply changes"]
    H --> I["audit log"]
    D --> J["conversion sync"]

Safety Rules

• Keep read and write paths separate.
• Require explicit approval before any spend-affecting write.
• Treat account binding, token storage, and billing as sensitive operations.
• Prefer dry-run or validation-only mode before real execution.
• Never let the model call the platform API directly without a deterministic adapter layer.

Common Roles

ads-router

• classify intent
• decide whether the request is connect, read, diagnose, propose, apply, or sync
• route to the correct platform adapter

ads-core

• own the normalized ad schema
• own the approval record schema
• own the audit and execution record schema

ads-read

• pull accounts, campaigns, ad groups, ads, creatives, budgets, reports, and conversion summaries
• never mutate

ads-diagnose

• identify wasted spend, fatigue, unstable CPA, budget constraints, and broken tracking
• output reasons, not commands

ads-propose

• produce a structured patch for review
• keep proposed changes small and explicit

ads-apply

• execute only approved changes
• keep a full result record

ads-conversion-sync

• send server-side events or conversion signals back to the platform
• treat this as an execution path with tracking and audit

Shared Output Contract

All platform-specific skills should be able to return at least:

• access_scope
• account_tree
• normalized_reports
• diagnosis
• proposed_patch
• approval_record
• execution_result
• audit_log

Platform Policy

Each platform adapter should declare:

• auth method
• readable objects
• writable objects
• conversion sync support
• expected rate-limit behavior
• known policy or access caveats

First Version Boundary

The first version should support:

• connect
• read
• normalize
• diagnose
• propose
• manual approve
• apply a small safe action set

The first version should not try to automate:

• full campaign generation from scratch
• unrestricted budget allocation
• multi-account spend migration
• self-directed creative replacement

Shared Product Selection Preferences

Shared routing rules for product-selection work.

Use this file first. It exists to help the agent choose the right path before collecting data or making recommendations.

Merchant Models

Classify the business model first:

• white-label: generic or lightly customized products, speed and margin first
• brand: differentiated products, trust and premium first
• distribution: authorized resale, assortment and channel leverage first
• content-led: creator or media-driven selling, content efficiency first
• hybrid: mixed model, choose the dominant constraint first

Do not confuse merchant model with sales channel.

Channels

Classify the main selling channel next:

• amazon: search-led marketplace
• tiktok-shop: content-led marketplace
• independent-site: audience and LTV-led storefront
• multi-channel: compare channel fit explicitly instead of averaging assumptions

Treat channel choice as a strategy question, not a formatting detail.

Question Types

Classify the user request before doing any work:

• platform data: listings, prices, reviews, rankings, creators, comments, shop pages
• social proof: cross-platform content demand, audience language, creator traction, comment signals
• supply chain: factory options, MOQ, lead time, customization, packaging, logistics feasibility
• financial judgment: margin, contribution profit, break-even CPA, inventory pressure, return sensitivity
• compliance risk: restricted claims, category restrictions, certification or safety exposure
• selection synthesis: whether a product fits a merchant model and channel combination

Do not answer a platform-data question from generic web search if a platform skill exists.

Skill Routing

Default route:

1. identify merchant model 2. identify channel 3. classify question type 4. collect the minimum valid evidence 5. synthesize only after evidence exists

Use these routes:

• Instagram/Meta demand or content proof -> social-media-extractor
• TikTok content and audience language -> tiktok-research
• Instagram creator/content/comment proof -> instagram-account-research or instagram-content-benchmark or instagram-audience-voice
• supply-side product or supplier data -> user-provided supplier sheets,

quotations, or approved marketplace exports

• cross-source sourcing judgment -> sourcing-selection
• Hook, structure, and why content works -> video-analysis
• Creative benchmark adaptation -> benchmark-to-brief

If no supply-chain or finance skill exists yet, state the missing layer clearly and avoid false confidence.

Minimum Judgment Order

Use this order unless the user asks otherwise:

1. Is there real demand or proof of demand? 2. What kind of competition is this: price, content, trust, supply, or search ranking? 3. Does this product fit the target channel? 4. Does this product fit the merchant model? 5. Can the unit economics survive ads, fees, shipping, and returns? 6. Is the compliance or operational risk acceptable?

Do not jump from "people are selling it" to "we should sell it."

Failure Patterns To Avoid

• mixing merchant model and channel into one label
• using TikTok signals to justify an Amazon-first launch without search proof
• using Amazon demand to justify TikTok Shop without content-demo fit
• using one social platform's proof as if it represents all audience demand without checking channel fit
• skipping margin and return sensitivity because top-line demand looks strong
• treating "product selection" as one workflow when the real bottleneck is supply chain or finance

Shared Public Skill Rules

Shared execution rules for released PostPlus skills running inside the PostPlus runtime.

Use this rulebook when a skill needs PostPlus Cloud service access, temporary local artifacts, or a compile step before provider execution.

PostPlus Cloud Rule

• Treat host-managed adapters, auth, and billing boundaries as the source of

truth for provider-backed execution.

• Do not probe local secret env vars or ask the user to paste provider tokens

into the chat just to see whether a capability exists.

• If the PostPlus Cloud boundary reports that a capability is unavailable, unauthorized,

or not configured, fail fast and report that directly to the user.

• On that capability-missing path, do not switch into a “we can collect later”

discovery flow.

Supported Script Rule

• Use PostPlus-supported scripts and PostPlus Cloud services as the supported

execution path.

• Do not replace PostPlus-supported scripts with ad hoc curl, fetch,

node -e, heredoc, or exploratory shell glue.

• If a supported script returns a stable proxy, DNS, network, or

infrastructure error, stop immediately and report that failure directly.

Parallel Request Rule

• When multiple tool calls, file reads, script requests, provider submissions,

or data-collection requests are independent, prepare their inputs first and dispatch them as a bounded parallel batch instead of running them one by one.

• Do not serialize independent requests just because they target different

files, accounts, URLs, keywords, assets, platforms, or scripts.

• Keep steps serial only when a later request depends on an earlier result, a

skill explicitly requires a serial queue, approval or quote confirmation is still missing, or the skill's own cost/rate-limit boundary requires smaller batches.

• For approved side-effecting or provider-backed work, submit approved

independent items concurrently within the skill's stated batch or concurrency limit. Do not create duplicate requests to mask slow or failing providers.

Async Provider Task Rule

• When a supported script returns pending, processing, generationHandle,

runHandle, providerUrls.get, or a collection-report.json, treat that as a real resumable checkpoint.

• Do not block the user's conversation by looping on provider polling when the

next useful action does not depend on the finished artifact.

• Tell the user the job is still running, name the durable checkpoint in

business terms, and continue with independent planning, review, or prep work when useful.

• Poll again only when the next step truly needs the finished result or when

the user explicitly asks to wait for completion.

• If there is no useful parallel work, run one bounded poll pass, report the

current status, and keep the resume command or checkpoint available. The resume command for a hosted media job is postplus media poll --handle <output.data.id>; for a public-content scrape (the s_... handle from research scrape) it is postplus research scrape --run-handle <runHandle>; for a hosted research collection it is postplus research collect --run-handle <runHandle>.

Work Folder Rule

• Temporary request files, actor-input files, raw datasets, index files, and

cache-like intermediates belong under the current work folder's .postplus/ directory.

• Keep final user-facing deliverables outside .postplus/.
• Treat .postplus/ as internal implementation state and do not explain its

management steps to the user unless the user explicitly asks for internals.

• Do not write runtime temp artifacts into installed skill directories.

Real File Rule

• When a tool or script expects --input <file.json> or a comparable file

argument, provide a real file path.

• If the input must be synthesized first, write it as a real file under

.postplus/ instead of passing inline JSON text.

Hosted Request Shape Rule

• Hosted skills run the converged verb grammar. The agent supplies only the

skill-specific input; the closed-source CLI runner translates it into the hosted request, mints identifiers, and derives billing dimensions.

• Each hosted command reads its input from --request <file>:

postplus research collect <collection-key> --request <input.json>, postplus research scrape <source-key> --request <input.json> (scrape input is a JSON array of { "url": ... } entries), postplus media <verb> <endpoint-key> --request <input.json>, postplus media analyze <model-key> --request <payload.json>, and postplus publish <operation> --request <input.json>. Flags-surface media endpoints take the input as --<flag> options instead and read no --request file.

• Before writing a --request file, read the public schema with

postplus research schema --collection-key <key> --json, postplus media schema --endpoint <endpoint-key> --json, or postplus publish schema --json.

• Put only the skill-specific request under --request (an object, or an array

for research scrape). Do not hand-write runner-managed fields such as ids, tokens, or billing dimensions; the CLI mints or derives them and rejects them in the request body.

• Pass shared execution fields as command-supported flags

(--quote-confirmation-token, --hosted-operation-id, or --run-handle to resume a pending research collection), not inside the skill-specific input.

• If a hosted command prints Quote confirmation challenge: <path>, run the

exact postplus quote confirm --json --challenge-file <path> command, then rerun the same hosted command with --quote-confirmation-token <token>.

Conversation Media Rule

• When the user shares images, videos, or other media inside a local AI agent

conversation, only save or upload the media if the agent can access a real local file path, attachment handle, clipboard bytes, or other actual binary source exposed by the host runtime.

• If the agent can only see the media as model context and cannot access the

original bytes, do not recreate, screenshot, summarize, or generate a substitute file and present it as the original asset.

• On that path, stop and tell the user the honest unblocker: provide a local

path, attach the file through a host mode that exposes a path, or save the media to disk before continuing.

• Do not send conversation media to PostPlus Cloud through inline base64 in the

hosted JSON request. Use the supported local file upload path when a real file or binary source exists.

Compile-Step Rule

• For complex collection families, compile the user brief into provider-ready

input before the expensive execution step.

• Inspect or adjust the compiled input when the request is high-cost, ambiguous,

or unusually broad.

Local Dependency Bootstrap Rule

• For approved local media dependencies in the current CLI-first release, the

user's agent must proactively install the missing dependency in the user's local environment before running the supported script that needs it.

• Current approved local media dependencies are:
• python3
• yt_dlp
• ffmpeg
• ffprobe
• Skill scripts must call the PostPlus CLI local dependency checks for

these dependencies. Individual skills must not hard-code OS-specific binary names, shell syntax, or install paths.

• The resolver owns platform command selection. It keeps macOS/Linux on the

canonical python3 path and uses the Windows Python launcher / Python 3 command candidates when the host platform is Windows.

• PostPlus CLI itself is not the installer for those tools.
• Do not ask a non-technical end user to install those tools manually or to

interpret tool names such as ffprobe.

• Use the smallest direct install path already supported by the host

environment.

• After installation, rerun a direct verification command before continuing.
• If installation or verification fails, stop immediately and report that

failure directly instead of inventing fallback glue.

Cost Discipline

• Default to a bounded first pass before a broader second pass.
• Treat the first pass as evidence gathering for inspection and iteration, not

as the full-market scrape.

Shared Research Preferences

Shared routing rules for research skills.

Core Rule

Classify the request first:

• Web facts: news, rules, company facts, public articles, policy context -> use web search
• Platform data: creators, hashtags, posts, comments, engagement, listings, rankings, reviews -> use the relevant platform skill first
• Content breakdown: hook, structure, shots, spoken lines, adaptation -> use analysis skill after collection
• Mixed: collect first, shortlist second, analyze third

If a platform skill exists, do not answer a platform-data question primarily from public web search. Use web only when:

• the user explicitly wants a fast public-web read
• the fact is external to the platform dataset
• no platform skill exists
• the platform skill path is blocked

Platform-First Rule

Treat these as platform-data requests by default:

• research data from a specific platform
• find bloggers / creators / KOCs / KOLs
• see which content is currently taking off
• inspect comments, engagement, price bands, rankings, and competitors
• find benchmark samples

Default execution order:

1. use the platform skill to collect a small valid dataset 2. normalize or rank locally 3. synthesize findings from the dataset 4. use deeper analysis skills only on shortlisted samples 5. use web search only as a secondary layer if needed

Preferred Routes

• Instagram and Meta social research routing -> social-media-extractor
• Instagram creator discovery -> instagram-creator-discovery
• Google search-intent and topic momentum -> google-trends-research
• TikTok platform data -> tiktok-research
• TikTok ad creative research -> tiktok-ad-research
• supply-side product or supplier data -> user-provided supplier sheets,

quotations, or approved marketplace exports

• Instagram creator discovery -> instagram-creator-discovery
• Instagram creator/account research -> instagram-account-research
• Instagram content benchmark -> instagram-content-benchmark
• Instagram audience language -> instagram-audience-voice
• Instagram hashtag/tagged campaign scouting -> instagram-campaign-scout
• YouTube channel and video research -> youtube-research
• Facebook page and post research -> facebook-research
• Hook, structure, and why content works -> video-analysis

Chaining Rule

Do not treat research skills as isolated.

Typical chain:

1. collect 2. normalize 3. rank or shortlist 4. analyze the strongest samples 5. adapt into strategy or creative output

If the user also wants execution after research, hand off only after the research output is explicit:

1. package into brief, shortlist, or publish-ready copy 2. confirm channels or destinations if execution is high-impact 3. then route to outreach or publishing skills

For Instagram creator discovery requests, add a routing step before collection:

1. normalize the brief 2. choose handle-first, content-first, graph-first, or mixed 3. collect with instagram-creator-discovery using a wider recall band when needed 4. enrich candidates 5. classify creator type 6. shortlist 7. hand off to outreach only after scoring

Clarify Only When It Changes The Route

If the user intent is broad and the next step is ambiguous, ask one short question.

Examples:

• Do you want to inspect an account shortlist first, or start with content benchmarks?
• Do you want a platform data scan, or a breakdown of specific content structures?
• Are you scanning the market first, or do you already have samples to analyze?

If the route is already clear, do not stop to ask.

Video Analysis Proactive Ask

When the user wants to analyze a specific video — hook, structure, shots, or why it works — ask if they want to use video-analysis before falling back to generic analysis. Skip the ask only if the user already named the skill.

Failure Pattern To Avoid

Common mistake:

• user asks a platform-specific strategic question
• agent interprets it as a generic strategy question
• agent answers from web search first

Correct response:

• recognize the named platform
• route to the relevant skill first
• collect the minimum valid evidence
• synthesize only after evidence exists

Creator-discovery-specific mistake:

• user asks for 5k-10k creators in a niche
• agent defaults to keyword-based account search
• follower filtering happens only after noisy recall

Better response:

• recognize that follower band + niche fit usually needs routing
• prefer content-first or graph-first when account search is likely to over-rank large accounts
• use instagram-creator-discovery for Instagram creator discovery
• use wider recall and tighter shortlist instead of strict first-pass filtering
• classify individual creator vs brand/product vs educator/consultant vs aggregator before producing outreach-ready leads

Cross-platform mistake:

• user asks a broad social question across multiple networks
• agent jumps straight into one familiar platform
• cross-platform differences are never surfaced

Better response:

• recognize when the request is about Instagram/Meta social direction rather than one named platform
• start with social-media-extractor
• use platform-specific skills only after the route is clear

Shared Source-Of-Truth Files

Use this reference when a skill needs stable campaign- or project-level context for repeated creative generation.

The goal is not to create more documents.

The goal is to stop every script, hook, storyboard, and render request from starting as a blank prompt with no memory.

Why This Matters

Short-form creative systems drift when they regenerate everything from one chat turn:

• brand voice shifts
• product claims become inconsistent
• personas lose credibility
• hooks repeat or flatten
• the next skill cannot tell what was grounded versus improvised

Stable source-of-truth files solve that.

Default File Set

Recommended default files:

• brand.md
• persona.md
• product.md
• hooks.md

These are shared campaign or project assets. They are not private implementation details of any one creative skill.

What Each File Owns

brand.md

Owns:

• brand voice rules
• forbidden phrases
• positioning
• message boundaries

Use it so the brief does not get lost in scattered chat threads or copied inconsistently.

persona.md

Owns:

• who is speaking
• why this person is credible
• relationship to the product
• repeatable delivery constraints

Use it to keep recurring AI UGC creators or narrator identities coherent across many assets.

product.md

Owns:

• product facts
• claims
• mechanism
• objections
• approved proof anchors

Use it so scripts, voice, and render layers do not invent unsupported product logic.

hooks.md

Owns:

• hook shells
• tested openers
• emotional triggers
• format variations
• anti-repetition memory

Use it as a living hook library. Do not automatically treat it as validated performance truth unless the surrounding workflow confirms that.

When To Recommend This System

Default recommendation signals:

• repeated script generation for one brand
• multiple products or SKUs
• multiple personas
• multiple platforms
• batch testing of hooks or angles
• need for continuity across script, storyboard, and render layers

For one-off exploration, this file system may be optional. For repeated campaign work, it should become the default bias.

Relationship To Skills

Creative skills may read these files.

Creative skills should not each reinvent their own local equivalent.

Typical pattern:

• shared or project setup creates or maintains the files
• script review consumes them for scripts
• storyboard and render skills consume the approved script objects downstream

User-Facing Framing

Explain this as a quality and consistency system, not as paperwork.

Good framing:

• "If you want stronger batch script quality, I recommend keeping brand, persona, product, and hook context in four stable files so later generations stay consistent."

Weak framing:

• "Please fill out these files before we can help."

Shared TikTok Music Workflow

Shared routing and chaining rules for TikTok music, sound, video sample, and local audio-reference work.

Core Rule

Classify the request by the object the user already has:

• No specific sound yet: find candidates first with tiktok-research
• Music URL / sound URL / musicId / song keyword: use tiktok-research

when the released TikTok collection path can produce sample video URLs; if it cannot, ask the user for selected video URLs or an existing sample dataset

• Selected video URLs or sample dataset: download videos and extract audio with tiktok-music-archive-downloader
• Local video or audio files: route through media-router into transcription, subtitles, or video-analysis

Do not start with downloading when the user has not selected a sound or sample set. Do not treat extracted audio as commercially cleared unless the user confirms rights or platform-licensed use. Download claims are limited to TikTok post URLs reachable from the user's current local browser/IP environment. If yt_dlp reports that the current IP is blocked, preserve the failed source URL and stop the archive/audio-extraction chain instead of retrying broad downloads or inventing an unapproved proxy/cookie path.

Skill Chain

Use this default chain for TikTok music research:

1. tiktok-research: discover candidate sounds by region, category, or campaign fit. 2. tiktok-research: collect or normalize selected video samples when the released collection path supports the request. 3. tiktok-music-archive-downloader: download representative videos and extract reference audio. 4. video-analysis: analyze video structure, hook, pacing, visual pattern, and usage context. 5. audio-transcription or video-transcription: transcribe lyrics, speech, or voiceover when needed. 6. subtitle-packager: produce SRT/ASS only after timed transcript artifacts exist.

Skip steps when the user already provides the corresponding artifact.

Collection Boundary

The released public surface does not expose a separate TikTok music-sound collector skill. Keep music-specific collection claims inside tiktok-research only when the released collection key and script path support the exact request. If they do not, ask for selected TikTok video URLs or an existing sample dataset before downloading.

Download Access Boundary

tiktok-music-archive-downloader is a local downloader for reachable TikTok post URLs. Passing the local dependency bootstrap does not prove that TikTok will let the current local IP/browser access path fetch the post. A download report item with failureCode: "tiktok_ip_blocked" is the supported fail-fast outcome: report the blocker with sourceUrl, stderr, and the report path, then stop until a reachable URL/sample file or approved access bootstrap exists.

Output Contracts

Music Candidate

Use for trend discovery outputs:

{
  "platform": "tiktok",
  "recordType": "musicCandidate",
  "musicId": "",
  "musicTitle": "",
  "musicAuthor": "",
  "musicUrl": "",
  "regionCode": "",
  "rank": null,
  "trendReason": "",
  "sampleVideoUrls": [],
  "sourceCollectionPath": "",
  "fetchedAt": ""
}

Music Video Sample

Use for sound collection outputs:

{
  "platform": "tiktok",
  "recordType": "musicVideoSample",
  "videoId": "",
  "videoUrl": "",
  "authorUsername": "",
  "text": "",
  "musicId": "",
  "musicTitle": "",
  "musicAuthor": "",
  "likeCount": null,
  "commentCount": null,
  "shareCount": null,
  "viewCount": null,
  "publishedAt": "",
  "sourceCollectionPath": ""
}

Download Manifest

Use for archive-download inputs:

{
  "items": [
    {
      "sourceId": "musicid-videoid",
      "sourceUrl": "https://www.tiktok.com/@user/video/123"
    }
  ]
}

Storage

Use this campaign layout:

customers/<customer-id>/campaigns/<campaign-id>/research/tiktok-music/
  raw/
  normalized/
  analysis/
  archive/<run-id>/
    manifest/
    videos/
    audio/
    index.json

Keep raw collection output, normalized datasets, and archive manifests. Do not keep only the final audio files.

Combination Cases

Use tiktok-research with this workflow when music fit must be judged against broader TikTok content, creator, hashtag, or comment context.

Use tiktok-ad-research separately when the request is about paid ads or Creative Center ads. Do not infer organic music trends from ad-only data unless the user asks for paid creative context.

Use video-analysis after downloading only the shortlisted strongest samples. Do not run semantic video analysis over broad unscreened trend results.

Use media-router when the user gives local files and the needed output is unclear. Let it choose transcription, subtitle packaging, semantic analysis, or edit prep.

Use creative-qa only after a human has reviewed candidate audio/video samples and wants structured review records.

Failure Patterns

• Jumping from a trend list directly to audio extraction without checking videos that use the sound.
• Ranking sounds only by trend rank instead of campaign fit and sample availability.
• Treating TikTok sound-page downloads as reliable when yt-dlp sound extraction may be broken.
• Retrying broad archive downloads after failureCode: "tiktok_ip_blocked"

instead of reporting the access blocker.

• Losing provenance by saving audio without source video URL, music id, and actor output.
• Presenting scraped or extracted audio as cleared for commercial reuse.

Shared User Guidance

Rules for proactive user communication when using PostPlus skills. Apply these before any skill-specific execution, especially when the user intent is broad or the interaction appears early in the session.

Orient Before Acting

When the user gives a broad request and hasn't named a specific skill, briefly name which skill will be used and what it does before proceeding. Then ask a short confirmation.

Keep it to one sentence. Don't run in silence.

Audience Language

Assume the user is a marketing operator, not an engineer. Use everyday business language in user-facing messages.

• Prefer "image file", "saved file", "upload", "account connection", "service

limit", and "next step" over implementation terms when the technical detail is not the real unblocker.

• Avoid unnecessary provider, API, schema, endpoint, payload, base64, runtime,

attachment handle, or storage internals in normal user-facing copy.

• If a technical term is the real blocker, translate it into what the user can

do next in the same sentence.

• Keep detailed technical evidence in local artifacts, logs, or task summaries

unless the user asks for implementation details.

Execution Expectation

Before running a skill that will collect data, call a PostPlus Cloud service, or write durable local artifacts, say what will happen in one compact sentence:

• the first action
• the expected output artifact
• the likely downstream handoff

Use the concrete skill names, not abstract workflow labels.

Good shape:

• "I will first use tiktok-research to collect a small public sample and output a shortlist; after approval, I will pass it to video-analysis for hook and shot breakdown."
• "I will first put the persona lock and source basis into the image-batch-runner request and output a local asset manifest; the next step can go to creative-qa or video-batch-runner."

Do not promise hosted, provider, file-reference, account-connection, or publishing behavior unless the current skill contract and registry release requirements already support that path.

Async Task Copy

When a long-running skill returns a pending generation, transcription, or collection checkpoint, tell the user two things in plain language:

• the job is running and can be checked again from the saved checkpoint
• the conversation does not need to pause while the provider finishes

Good shapes:

• "The render is running now. I saved the checkpoint, so I can keep preparing the QA checklist and check the render again when we need it."
• "The collection is still processing. I can continue outlining the summary structure now, then poll the saved report before filling in metrics."

Do not keep the chat idle just to poll. Poll immediately only when the user's next requested step depends on the completed artifact or the user explicitly asked to wait.

Technical Boundary Orientation

Before an expensive collection, media generation, transcription, publishing, or large local processing step, inspect the skill's own boundary section and public metadata requirements. If the request crosses a known limit, do one of these before execution:

• internalize the boundary into the request shape, such as bounded first passes,

segment contracts, capped frame counts, or compiled hosted inputs

• ask one short scope question when the skill cannot choose safely
• stop with a direct blocker when the released contract does not support the

requested path

Do not ask the user to understand provider limits, endpoint names, actor fields, polling internals, or local implementation details unless that detail is the real unblocker. Use the business meaning of the limit instead.

Brief Decomposition

When the user gives a fuzzy business goal ("help me promote this product", "make a viral video"), do not jump into execution. Decompose the goal into 2-4 concrete sub-tasks, map each to a PostPlus skill, and present as a numbered plan. Then ask which step to start with.

Common decomposition templates:

Use these as reference, not a rigid checklist. Adapt to the user's actual scope.

Cross-Skill Suggestion

After completing a skill, offer one concrete downstream skill as the logical next step. One suggestion at a time. Don't list all possibilities.

Examples:

• tiktok-research done → "Want me to break down the hooks and structure of these videos with video-analysis?"
• video-analysis done → "These insights could feed into a campaign brief via benchmark-to-brief."
• benchmark-to-brief done → "Brief is ready. Want me to turn it into a video request plan?"

First-Use Mini Onboarding

If this appears to be the user's first time using a skill, add one line of context before executing. Explain what the skill does and when it's typically used. Don't over-explain — one sentence is enough.

Boundary Redirect

If the user asks for something this skill does not handle, name the right skill immediately instead of saying "I can't do that" or attempting a fallback.

Failure Copy

When a supported script or PostPlus Cloud service fails with a stable error, stop and report the blocker directly. The user-facing failure message must include:

• the skill and script or PostPlus Cloud service that failed
• the exact boundary that blocked execution
• the missing input, dependency, account connection, hosted endpoint, or file

contract when known

• the next honest unblocker

Do not say a task is queued, partially completed, or recoverable unless a real artifact exists and the skill contract defines that recovery path.

For images or other media pasted into a local AI agent conversation, be explicit about whether the agent has access to the original asset bytes or only visual model context. If the host runtime does not expose a readable path, attachment handle, clipboard bytes, or binary source, do not claim that the original media was saved or uploaded. Tell the user to provide a local path or save the asset to disk before continuing.

Good shapes:

• "video-analysis stopped before provider analysis: the upload reached Gemini Files API, but the file never became ACTIVE within the hosted wait window. The real unblocker is provider file activation, not another local retry."
• "image-batch-runner cannot use image-gpt-image-2-text in this release: the skill runner and registry only expose hosted endpoints already present in the PostPlus media-generation catalog."
• "social-media-publisher preview succeeded, but publishing is still approval-gated; no post was sent."
• "I can see the pasted image in this conversation, but this Codex session has not exposed the original image bytes or a local file path to tools. Please save the image locally or provide its path, then I can upload that real file."

Continuity Copy

When a generation request claims continuity across segments, say whether that continuity is text-only or actually bound by image or audio evidence.

Prefer short, honest wording:

• "人物连续性这次还是文字约束；如果你愿意补一张人物图，会更稳。"
• "商品和 App UI 已有图片绑定，这部分会比人物更稳。"
• "声音这次还没锁；如果你在意角色连续性，我建议先固定一个 voice take。"
• "这版可以直接跑，但更准确地说是已约束，不是已锁定。"

Do not tell the user continuity is locked, confirmed, or guaranteed when the request is only text-constrained.

Keep It Brief

All of the above must fit in 1-2 sentences per interaction point. Proactive communication builds trust; verbose communication erodes it. If you have more to say, wait for the user to ask.