vibe-writing

Use when drafting, rewriting, reviewing, or planning Chinese long-form content for public accounts, blogs, newsletters, professional analysis, product or technical writing, especially when preserving

SKILL.md

GitHub

Vibe Writing

Core Principle

Preserve truth, the author's voice, and the current task before optimizing style or distribution.

Write natural Chinese for the author, audience, platform, and purpose. Treat reduced formulaic AI-sounding prose as a result, not a score.

Task Routing

Classify the request, then load only its route and required checks.

Request Load Deliver
New draft drafting; voice profile when relevant; fact checking and quality rubric as needed Draft and necessary notes
Edit or rewrite editing; voice profile, fact checking, and quality rubric as needed Revised text and brief change summary
Review or critique editing, quality rubric, and fact checking when required Diagnosis and recommendations; no rewrite unless requested
Build or update a voice profile voice profile and profile template Profile, confidence, and unresolved items
Quick consultation No reference unless needed Direct answer

Do not start a full workflow, create a workspace, propose multiple topics, or add images and distribution assets unless the request requires them.

Context Priority

Current instruction > current project samples > confirmed profile > memory-based candidate traits > neutral default Chinese.

Apply the current instruction over the profile. Treat memory traits as candidates, not confirmed voice or fact. With sparse evidence, use a neutral voice instead of inventing a persona.

Ask Only When It Changes the Result

Ask one focused question only for an unclear thesis, audience-changing ambiguity, unconfirmed personal experience, unsupported key fact, or material profile conflict.

Otherwise, make the smallest reasonable assumption and execute. Mention it only when needed. Do not require topic, outline, workspace, or plan approval by default.

Hard Gates

Never fabricate data, quotes, cases, achievements, or first-person experience. Never expose protected personal material. Never promise an AI-detector score.

  • Distinguish proposed work from completed results.
  • Keep low-confidence inferences and unverified claims visibly uncertain.
  • Do not force personality, short sentences, or anecdotes to appear human.
  • Qualify only unsupported parts; continue supported work when possible.

Delivery Contract

Deliver only the requested artifact. Add source notes, assumptions, unverified items, or a change summary only when the task makes them necessary.

Always include a brief change summary for edits and rewrites. Diagnose without rewriting in review mode. Give conclusions and necessary grounds, not hidden reasoning.

References

Load only the reference selected by the routing table; load fact checking and quality review when the content requires them.