Readmedaddy

readmedaddy

Generate the best possible README for the work in front of you — a "Fable-class" README — by reading what the project actually is, borrowing the conventions of the best READMEs for that kind of project, drafting candidate sections, and ranking them through a contextual multi-gate rubric until the front page earns its first screen.

What "Fable-class" means here. A README that, in the first screen, tells a stranger what this is and why they'd care; gets them to a working first result in under ~30 seconds; stays skimmable; is honest about limits; and follows the conventions a reader of this archetype already expects. Not a history lesson, not a feature dump, not AI slop. Front-load value, prove it, then let the curious scroll.

Core principle. A README is judged by its archetype. A CLI lives or dies on a demo and a one-line install; a library on copy-pasteable API usage and trust badges; a data/ML project on results and citations; an agent skill on what triggers it and a worked example. The same section that's essential for one is bloat for another. So detect the archetype first, weight the rubric to it, and never apply a generic template. Match the work to its peers, not to a checklist.

Run this first: detect the ARCHETYPE

Before drafting a single line, classify the project. Read the repo, don't guess: check the manifest and entrypoints (package.json bin vs main/exports, pyproject, Cargo.toml [[bin]] vs [lib], SKILL.md, Dockerfile/compose, notebooks, go.mod), the directory shape, and any existing docs. Assign one primary archetype (a project may be a blend — pick the dominant reader's expectation):

• CLI — installed and run from a terminal. Reader wants: a demo (gif/asciinema) and the smallest real command.
• library / package — imported into other code. Reader wants: install + minimal API usage, badges, API surface.
• framework — scaffolds or hosts other code. Reader wants: mental model, quickstart, conventions, when-to-use.
• app / SaaS — a running product or service. Reader wants: what it does, a screenshot, how to deploy/run, status.
• infra — deploys or operates systems. Reader wants: prerequisites, topology, apply/teardown, safety/secrets.
• data / ML — models, datasets, experiments. Reader wants: the result up front, benchmarks/tables, reproduce steps, citation.
• agent-skill / agent-tool — a Claude Code skill, MCP server, or LLM tool. Reader wants: what it does, when it triggers, and a worked example.
• research — a paper, study, or finding. Reader wants: abstract/result, figures, citation, honest scope.
• monorepo — many of the above. Reader wants: a map of packages and where to go.
• internal tool — audience is teammates, not the public. Reader wants: why it exists, ownership, a runbook (run / deploy / rollback), gotchas.

State the detected archetype in one line and proceed. If genuinely ambiguous or the user contradicts you, ask once; otherwise assume and note it. Load the archetype's profile from references/archetypes.md for its section menu; the gate weights live in references/multi-gate-rubric.md.

Then: generate → multi-gate-rank → iterate

1. Pull the patterns. Open references/famous-readme-patterns.md for the conventions and exemplars of the detected archetype — section order, what the first screen does, what earns its space, what to omit.
2. Generate candidates. Draft the README, and where a section has real choices (the hook line, the quickstart command, whether a diagram earns its place), generate 2–3 genuinely distinct candidates rather than one safe default. Pull every fact — commands, install lines, names, claims — from the repo, never from assumption. If you can't verify it, don't assert it.
3. Rank through the gates. Score the draft against the ten gates below (0–5 each), weighted by archetype per references/multi-gate-rubric.md. Keep the highest-scoring candidate per section; rewrite anything scoring ≤2 on a heavily-weighted gate.
4. Iterate until it clears. Re-rank after each rewrite. Stop when the heavily-weighted gates for this archetype are strong and nothing material is missing — not when a generic template is "filled in".

The full generate→rank→merge workflow — including when to escalate to a full tournament of whole-README drafts with a judge panel — is in references/generation-and-ranking.md.

Run the draft through the G10 anti-slop tells in references/multi-gate-rubric.md before finishing: kill "in today's fast-paced world", negative-parallelism stacking ("it's not X, it's Y" repeated), gratuitous rule-of-three, and empty AI vocabulary. Voice is a gate (G10), not a finishing polish.

The multi-gate README ranking rubric

Score each 0–5. Weights are contextual to the archetype — see references/multi-gate-rubric.md for the per-archetype weight table.

• G1 Hook — the first screen instantly conveys what it is and why you'd care: a one-line value prop, not a backstory.
• G2 Identity & trust — name, one-liner, badges (CI / version / license / downloads), tasteful social proof.
• G3 Visual — logo / ASCII / diagram / screenshot / gif that is apt and earns its space, never gratuitous.
• G4 Quickstart — install plus the smallest real usage in under ~30s, copy-pasteable and correct.
• G5 Scannability — heading hierarchy, short paragraphs, tables, a TOC when long; skimmable in seconds.
• G6 Completeness — usage, config, examples, links, contributing, license — without bloat.
• G7 Credibility — real examples, tests/CI signals, honest limitations and non-goals.
• G8 Contextual fit — follows the conventions of its archetype (CLI vs library vs app vs infra vs data/ML vs skill vs research).
• G9 Community & maintenance — contributing, code of conduct, changelog, support/roadmap signals.
• G10 Voice — distinct, confident, no AI slop (no "fast-paced world", no negative-parallelism stacking, no rule-of-three padding).

Archetype weighting (examples). CLI → G3 (demo) + G4 heavy. Library → G4 (API usage) + G2 (badges) + G6. App/SaaS → G1 + G3 (screenshot) + G9. Data/ML → G1 + G7 (results/benchmarks) + citations. Agent-skill → G1 + triggers + G4 (examples) + G8. Research → abstract/results + citations + G7. The full table is in references/multi-gate-rubric.md.

Composition and precedence

This skill decides README content and shape. It yields, highest first, to: host safety policy; an explicit live user instruction; standing project docs (CLAUDE.md / CANON.md / AGENTS.md) and any house README template; then this skill's defaults. If the project mandates a README structure, follow it and apply the gates within it. Never invent facts, results, badges, or links the repo doesn't support — an honest thin README beats an impressive fictional one.

Operate offline. Every fact comes from the local repository; never fetch anything over the network to write a README — no web searches, no remote templates, no URL lookups. The repo in front of you is the entire input. (The patterns this skill draws on are baked into its reference files for exactly this reason.)

Keep it fresh (auto-update hook)

A README rots the moment the code moves on without it. An optional Claude Code Stop hook watches a project's signal files — manifests, entrypoints, CI, install scripts — and, when they change while the README does not, prompts you to refresh it through this skill, in the same session. It detects drift and asks; it never rewrites files on its own and never breaks a session. install.sh wires it up in one step. Full behavior, config, and modes are in references/auto-update-hook.md.

init — guard a repo

When the user says "readmedaddy init", "guard this repo", or "set up readmedaddy": configure the whole system through the wizard script — it is the single serializer; never write .readmedaddy.json or the workflow YAML directly.

1. Detect first, silently: git rev-parse --show-toplevel, README path via git ls-files, default branch, whether origin is GitHub, existing .readmedaddy.json (reconfigure mode).
2. Ask at most three questions, one at a time, stating every detected fact as an assumption the user can veto: nudge mode (auto/notify/enforce/off), CI gate (comment/fail/off — comment is the advisory-first default), and watch-list confirmation.
3. Two-step through the script: run python3 scripts/readmedaddy-init.py --print with every answer as a flag (--mode --pr --main --sweep --watch --readme --badge/--no-badge --no-hook), show the user the output as the preview, then on approval re-run the same flags without --print to write. --yes = the recommended preset.

The wizard makes zero network calls; it writes only the config, the workflow file, one badge line, and (optionally) the Claude Code hook registration — each enumerated in the preview.

What's in references/

• references/archetypes.md — the ten archetypes: detection signals, the section menu each reader expects, and what to omit.
• references/famous-readme-patterns.md — per-archetype exemplars and the patterns that make the best READMEs work.
• references/multi-gate-rubric.md — the ten gates with 0/5 anchors, the per-archetype weight table, the G10 anti-slop tells, and the merge procedure.
• references/generation-and-ranking.md — the generate→rank→iterate workflow and the whole-README tournament for high-stakes front pages.
• references/auto-update-hook.md — the readme-drift Stop hook: how it detects drift, its modes (auto / notify / enforce), config, and one-step install.