grounded-writing — Claude Code skill: what it does, how to install
A skill for drafting and reviewing developer docs, reports, and replies without invented facts, promotional filler, or irrelevant implementation mechanics.
npx skills add chrisbanes/skills --skill grounded-writing -a claude-codeWhat grounded-writing does
This skill establishes an editorial workflow based on Chris Banes's writing style. It keeps reasoning visible and requires every claim to rest on supplied evidence or verified public facts. The agent must never invent personal experiences, benchmark numbers, or subjective opinions. If a missing personal position would alter the argument, the model pauses and asks.
For public developer documentation, the skill strips internal mechanics and highlights observable differences and API contracts. Diagnostic details remain only in internal design documents and technical reports where they serve a direct purpose. Reviewing happens in two distinct passes: first for truth and material accuracy, then for tone and style.
Drafting prose in the chat output does not authorize file edits in the repository.
What it helps with
- Review public developer documentation without leaking implementation details
- Strip promotional framing and unsupported certainty from drafts
- Check internal technical reports for factual claims and material errors
- Draft concise replies to technical pull request reviews
Example prompts for the agent
- Review this draft of our new API documentation and flag any unsupported claims.
- Edit this architecture decision record using grounded-writing principles.
- Draft a concise reply to this pull request review comment without promotional filler.
How to install grounded-writing
1Skills CLI (Vercel)
npx skills add chrisbanes/skills --skill grounded-writing -a claude-code2Manually: put the files in place
~/.claude/skills/grounded-writing/SKILL.mdCopy the whole skill folder; for one project use .claude/skills/grounded-writing/
1Skills CLI (Vercel)
npx skills add chrisbanes/skills --skill grounded-writing -a codex2Manually: put the files in place
.agents/skills/grounded-writing/SKILL.mdFor all projects use ~/.agents/skills/
1Skills CLI (Vercel)
npx skills add chrisbanes/skills --skill grounded-writing -a cursor2Manually: put the files in place
.cursor/skills/grounded-writing/SKILL.mdCursor also reads .agents/skills/ and .claude/skills/
1GitHub CLI
gh skill install chrisbanes/skills grounded-writing2Manually: put the files in place
.github/skills/grounded-writing/SKILL.md1Skills CLI (Vercel)
npx skills add chrisbanes/skills --skill grounded-writing -a gemini-cli2Platform CLI
gemini skills install https://github.com/chrisbanes/skills --path skills/grounded-writingManually: put the files in place
.devin/skills/grounded-writing/SKILL.mdThe legacy .windsurf/skills/ path still works
Skills CLI (Vercel)
npx skills add chrisbanes/skills --skill grounded-writing -a opencodeClaude.ai
Zip the folder and upload it in Settings → Capabilities → Skills
Step by step, with commands for each platform: How to install skills in Claude Code: five ways
Questions about grounded-writing
How do I install grounded-writing in Claude Code?
Run npx skills add chrisbanes/skills --skill grounded-writing -a claude-code. You can also place the skill definition file at ~/.claude/skills/grounded-writing/SKILL.md.
What does the grounded-writing skill do?
It guides the assistant through drafting and reviewing public docs, internal reports, and review replies. The procedure enforces evidence-based claims, cuts marketing filler, and removes internal mechanics that readers do not need.
How does it handle unverified claims or missing author context?
The skill forbids inventing opinions or results. If a missing detail materially affects the argument, the model pauses to ask rather than smoothing over the gap.
Source text
SKILL.md · Apache-2.0
--- name: grounded-writing description: Use when drafting or reviewing public developer documentation, other text the user will publish or send, or a user-owned draft under an explicit editorial review, including review comments, replies, and internal technical reports. --- # Grounded Writing ## Core principle Use Chris Banes's writing as the sole style model. Make the reasoning visible at the scale the artifact supports. Preserve the user's meaning and evidence; remove anything invented, generic, or included only to imitate a personality. ## Procedure 1. Establish scope and authority. Apply this skill to text the user will publish or send, at any length, or to an explicitly requested editorial review of their own draft. Identify whether the task permits drafting, editing, or recommendations only. Preserve the audience, purpose, and requested format. Exclude ordinary assistant replies, quoted source text, and prose attributed to someone else. For internal report reviews, apply truth and material-edit checks without imposing public-documentation style. 2. Establish the evidence and the user's actual position. Never invent personal opinions, experiences, results, preferences, memories, or certainty. Research discoverable public facts when research is in scope. If a missing personal position would materially change the text, ask and pause that part. Otherwise qualify uncertainty or mark a necessary gap conspicuously; do not add placeholders for details that do not affect the artifact's intended use. 3. Select detail for the reader. When public developer documentation compares options, check the observable difference, practical trade-off, and when to choose each separately. Preserve supported public contracts, shared guarantees (including compatibility), meaningful limitations, and context needed to interpret claims. Include implementation mechanics only when required for correct API use, necessary to distinguish a reader-facing choice, or explicitly requested. In reviews, explicitly recommend removing each irrelevant mechanic and name missing decision information rather than inventing it or asking for more internals. Distinguish behavior from implementation: preserve a supported return to another setting even if its timing is unverified; omit the internal timing term and delay unless their public or decision relevance is established. Verify uncertain contracts or relevance when research is in scope; otherwise flag material gaps rather than silently dropping or inventing behavior. Internal reports and design documents may need implementation and diagnostic detail; retain it for their intended use. 4. Read [the style profile](references/style-profile.md) and shape the argument before polishing. Use only the reasoning the artifact needs: a concrete observation, explanation, evidence, relevant limit, and practical consequence. A short reply may need just the actionable point and one supporting fact. 5. Review once for truth and once for style, including the headline and opening claim. Remove or qualify promotional framing and unsupported certainty. For material-only edits, identify the false claim or specific misreading that would prevent the intended use before changing anything. Optional polish is not a material fix. Add reproducibility details only when a stated claim or action depends on them; an internal report need not become a rerun protocol. Leave an adequate draft unchanged and briefly explain the relevant evidence, mechanism, or qualification that supports that conclusion. 6. Deliver the authorized result: write the requested text for a drafting task, apply fixes for an editing task, or give concrete recommendations for a review-only task. Drafting in the response does not authorize file edits. Resolve failed checks below within that boundary. If resolution needs an unknown personal position, ask rather than smoothing over the gap. ## Finish gate Finish when the authorized draft, edit, or review: - Preserves the requested purpose, format, and material-edit boundary. - Grounds every personal claim and substantive fact in supplied or verified evidence, honest qualification, or a clearly identified necessary gap. - Meets the reader-detail checks in step 3, retaining supported behavior and guarantees without irrelevant mechanics. - Uses the appropriate register and language, with concrete reasoning, useful caveats, and no promotional filler or repeated conclusion.
© chrisbanes · Apache-2.0 · chrisbanes/skills