Project Readme

Project README

Use this skill to build README files from repository facts. README structure is selected by the primary reader and their first successful action, not by language ecosystem alone.

Workflow

1. Classify the request:

   • create: no README or user asks for a new README.
   • rewrite: user asks for a full rewrite or replacement.
   • update/improve: user asks to improve, refresh, polish, fix, or update an existing README or package metadata.
   • validate: user asks only to check, audit, review, or validate.

2. Inspect the repository before writing:

   SCRIPT=$(find ~ -maxdepth 8 -name "inspect.py" -path "*/project-readme/scripts/*" 2>/dev/null | head -1)
   [ -n "$SCRIPT" ] && python3 "$SCRIPT" . || echo "Script not found; inspect manifests and repo facts manually."
   

3. Review the inspector output, then make the final README type decision:

   • ecosystem controls install commands, package metadata, badges, dev commands, and language-specific examples.
   • readme_type is a script-generated starting point, not final authority. Confirm it against the primary reader, first successful action, public API, source layout, examples, and package metadata before choosing the structure.
   • readme_type_candidates is the ambiguity set. If candidates disagree or the first successful action points elsewhere, choose the type best supported by repo facts and briefly note the override in your working notes. Ask one clarifying question only when repo facts cannot resolve the reader journey.
   • agent_skills lists detected skill files and plugin manifests when readme_type is agent-skill; use it to build collection indexes and validation sections.
   • media_assets lists detected logo, icon, banner, or screenshot candidates; use them only when they clearly represent the project.
   • readme_variants lists sibling README variants for multilingual workflows.
   • If classification affects the whole README structure and remains ambiguous, use a fresh classification pass or subagent when available; otherwise do a separate manual pass before writing.

4. Load only the needed modules:

   • Ecosystem: ecosystems/<ecosystem>.md, or each ecosystem when ecosystem is mixed.
   • Type: project-types/<readme_type>.md.
   • Shared ordering: references/section-order.md.
   • Badges: references/badge-patterns.md.
   • Validation rules: references/validation.md.
   • Writing improvement: references/writing-improvement.md for update/improve, rewrite, or README quality requests.
   • Research pass: references/research-pass.md for every rewrite and for stale/incomplete README improvement work.
   • Positioning: references/positioning.md when the user asks for comparison, repo docs name alternatives, or public incumbents are obvious.
   • Reader context: references/reader-context.md when audience context matters or the README is OSS, personal, internal, or config-oriented.
   • Markdown mechanics: references/markdown-mechanics.md when changing links, anchors, tables, alerts, code fences, images, or README variants.
   • Section modules in sections/ only for sections the README will include.
   • License section: sections/license.md when the README includes license terms or copyright credentials.

5. For rewrite, run a project research pass before drafting. Treat the old README as context only: it may be stale, incomplete, wrong, or missing important capabilities. Derive claims from manifests, exports, source files, examples, tests, docs, workflows, releases, package metadata, and other repository evidence.

6. For update/improve, diagnose before editing. Read the existing README and any in-scope metadata such as package.json, then identify concrete issues in reader fit, first-screen clarity, stale claims, examples/API usefulness, navigation, and metadata alignment. Make targeted edits based on that diagnosis; do not treat validation output as the improvement plan.

7. Compose from facts. Preserve true content from source files, existing README, docs, examples, tests, manifests, workflows, license files, and env examples. Do not invent package names, APIs, badges, deploy URLs, metrics, model performance, dataset size, support contacts, registry status, or roadmap claims.

8. Keep the first screen short. It must answer:

   • What is this?
   • Who is it for?
   • What is the fastest real thing the reader can do?

9. Validate before finishing:

   SCRIPT=$(find ~ -maxdepth 8 -name "validate.py" -path "*/project-readme/scripts/*" 2>/dev/null | head -1)
   if [ -n "$SCRIPT" ]; then
     python3 "$SCRIPT" README.md --repo .
   else
     echo "Validator not found; check manually against references/validation.md."
   fi
   

Core Rules

• Do not treat language ecosystem as README type. A Node repo may be a library, CLI, web app, GitHub Action, VS Code extension, MCP server, docs site, template, or monorepo.
• Dockerfile is a deployment signal unless the container image is clearly the product.
• Do not let a keyword override the reader journey. A library used by MCP servers is lib-or-sdk, not mcp-server, unless the repo actually implements and runs an MCP server.
• Badges are emitted only when the corresponding fact is real.
• Development commands must match the detected package manager and existing scripts; public registry install examples should default to correct alternatives for common package managers.
• Code examples must match real exports, routes, commands, tools, charts, modules, datasets, models, or files.
• License sections should include author or copyright-holder credentials when the license file or package metadata proves them; otherwise link the license plainly and do not invent ownership.
• README rewrites always require internal project research before drafting; never assume the old README is accurate or complete.
• Incumbent comparisons must use cited external sources or repo-named alternatives; omit comparison sections when they would force guesswork.
• Multilingual README edits must preserve code, commands, paths, links, badges, and selectors unless the user explicitly asks to rewrite them.
• If a section is unsupported by repo facts, omit it or label it as a manual step the user must supply.
• Validation is the final truthfulness and mechanics check. It does not replace writing analysis for improvement requests.

README Types

Supported readme_type values:

lib-or-sdk, cli, webapp, api-service, mcp-server, agent-skill, monorepo, docs-site, github-action, extension-plugin, container-image, terraform-module, helm-chart, model-card, dataset-card, research-code, template-starter, example-demo, spec-protocol, desktop-mobile-app, unknown.

If unknown, create a conservative repository README: state observed facts, provide verified setup commands only, and avoid API, deployment, registry, or performance claims unless source files prove them.

Source: antonio-orionus/skills — distributed by TomeVault.