| name | unslop-docs |
| description | Bans AI-slop documentation and rewrites README/docs into plain, verifiable guidance for humans. Removes emoji headers, badge walls, labeled-bullet Features, marketing fluff, install/API claims that don't match the repo, changelog empty phrases, unfit CONTRIBUTING boilerplate, and prompt/chat leakage. Use when the user says "unslop-docs", "AI README", "docs slop", or asks to fix documentation that looks generated, vague, or inaccurate. |
Unslop Docs
AI docs fail as generator chrome (emoji headers, badge walls, skeleton sections) and as unverifiable fiction (install commands, APIs, and features that don’t match the repo). This skill bans both and rewrites until docs teach a human what the project is, how to run it, and what will go wrong — with claims checked against the tree.
Core Mandate
- Verifiability over polish. Every install command, flag, endpoint, and public symbol must exist in the repo (or generated API truth). Wrong docs are bugs.
- Identity first. Line 1 answers “what is this?” before benefits.
- Ban generator fingerprints. No emoji headings, badge walls, labeled-bullet Feature pages, or stock skeletons filled with fluff.
- Prefer delete. Marketing adjectives, restated signatures, and boilerplate that doesn’t fit this repo go.
- Rewrite from notes, don’t synonym-humanize. Treat AI drafts as scrap; write plain docs from verified facts.
- Pivot rule. If a doc hits 3+ banned patterns, rewrite the page — do not emoji-strip and keep the same empty outline.
Hard Bans (never ship)
| Cluster | Banned |
|---|
| Chrome | Emoji in headings; badge walls (>~5 or non-resolving/placeholder shields); centered logo+emoji TOC as the first screen; --- before every heading |
| Structure | Skeleton-only Overview→Features→Install→Usage→Contributing→License with generic filler; labeled bullets (**Zero-config.** …) as the default Features shape; rule-of-three stacks as substance; rhetorical H2s (“Pick your path”, “What you get”) |
| Claims | seamlessly / empower / robust / cutting-edge / enterprise-grade / modern best practices; Features that are only stack names; soft install (“ensure Node is installed”) without pins |
| Fiction | Install/usage that doesn’t match manifests/CI; invented CLI flags/APIs/endpoints; Mermaid/architecture that doesn’t match modules; changelogs of “various improvements” / “enhanced UX” |
| Leakage | Certainly!, I hope this helps, turn0search, oaicite, utm_source=chatgpt, [Your Name], INSERT_…; HTML comments / invisible prompts with agent instructions |
| Boilerplate | Stock CONTRIBUTING/CoC with wrong project name, contacts, or process |
Full catalog: references/banned-patterns.md.
Workflow
1. Scope
- Full unslop: README,
docs/, API docs, CHANGELOG, CONTRIBUTING/CoC that ship with the project.
- Targeted unslop: named files or doc sections only.
Do not invent product claims. If truth is unknown, say so or omit.
2. Audit
Grep for emoji headings, leak tokens, marketing fluff, and shields. Dry-run install commands against package manifests and CI. Check every documented API/CLI symbol against source or generated truth.
Critical: hallucinated commands/APIs, prompt leakage, non-resolving badges that imply false status.
Use references/review-checklist.md.
3. Lock a docs contract
Before rewriting:
- One-sentence identity (“X is a …”)
- Primary audience and the one happy-path setup
- Source of truth for APIs (OpenAPI,
--help, typed exports, examples in CI)
- What not to document (internal, unstable, wrong for this repo)
- Plain markdown rules (no emoji headings unless house style requires)
Antidotes: references/craft-antidotes.md.
4. Rework
- Strip chrome and fluff; open with identity.
- Features = user-visible capabilities with one concrete example each — not dependency lists.
- Install: one verified path with pinned versions; copy commands from CI or a smoke script when possible.
- Teach: concept → minimal example → failure modes → link to reference. Cut signature restatements.
- Changelog: user-visible deltas; breaking changes first; no empty benefit-washing.
- CONTRIBUTING/CoC: only keep what matches real process and contacts; delete the rest.
- Remove leak tokens and agent-instruction HTML comments from published docs.
5. Verify
Re-run the checklist. Prefer running documented commands when the environment allows. Fail the skill completion if install/API claims were not checked.
Completion Report
- Scope audited
- Unverifiable claims fixed or removed
- Fingerprints removed (by cluster)
- Docs contract (identity + happy path)
- What was not verified and why
- Checklist status
Reference Routing