readme-doctor
Audit a GitHub repo's README against best-practice patterns and produce a prioritized punch list of fixes. Runs a structured review covering hero presence, install-to-first-success length, "what is this in one sentence" clarity, audience-jargon match, scannability, and drift signals (stale versions, dead links, badge sprawl). Read-only diagnostic; opens a PR only when the user explicitly asks.
livlign · 已有静态检查记录
安装与使用
运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。
npx skills add livlign/claude-skills --skill readme-doctor技能说明
name: readme-doctor description: Audit a GitHub repo's README against best-practice patterns and produce a prioritized punch list of fixes. Runs a structured review covering hero presence, install-to-first-success length, "what is this in one sentence" clarity, audience-jargon match, scannability, and drift signals (stale versions, dead links, badge sprawl). Read-only diagnostic; opens a PR only when the user explicitly asks.
readme-doctor
Most README problems aren't typos or missing sections — they're shape problems: the hero never lands, install-to-first-success buries the lede, the "what is this" sentence assumes prior context, the badge row is louder than the content. This skill audits a README against patterns that consistently correlate with maintainer outcomes (stars, contributor onboarding, issue quality) and produces a punch list a maintainer can act on in an afternoon.
The skill's quality comes from the rubric being grounded in the repo's actual audience and category, not from a generic checklist. A README for a 50K-star framework needs different things than a personal-project utility.
Phases
- Discovery — pick operating mode (Auto / Semi-auto / Manual, §1.1), input the README, infer the repo's category and audience from the scan, agree on what "good" looks like for this specific repo.
- Audit — walk the rubric, score each criterion, gather evidence (cite line numbers).
- Punch list — prioritize findings P0 (blocking) / P1 (high-leverage) / P2 (nice-to-have), each with a one-line fix.
- Optional output — read-only summary by default; on user request, open a PR with concrete edits.
Phase 1 — Discovery
1.1 Operating mode (ask first)
Same three modes as repo-visuals — Auto, Semi-auto (recommended), Manual. Use AskUserQuestion. Mode affects how many decisions are silent vs surfaced; it does not skip rubric checks.
1.2 Input
User may provide:
- GitHub URL →
gh repo view+ clone shallow - Local path → read directly
- Pasted README text → analyze in place, but the rubric loses any check that needs file-tree context (manifest signals, screenshot presence,
examples/dir, etc.). Flag these as "skipped — paste-only mode." - Nothing → ask first
1.3 Scan (collect before judging)
- README full text + rendered length (lines, words, time-to-scan estimate at 250wpm)
- Manifest:
package.json,Cargo.toml,pyproject.toml, etc. — version, description, keywords - File tree (depth 2): does an
examples/exist?docs/? Image assets? - Recent commits (last 10): is the repo active? When was last release?
- GitHub signals: stars, topics, open issues count, latest release date
1.4 Repo classification (drives rubric weighting)
Categorize the repo from the scan — different categories get different rubric weights:
- Library / package — installable, imported by other code (npm, PyPI, crates). Weight: install-to-first-success, concrete example, API stability signal.
- CLI tool — terminal-run binary or script. Weight: one-line install, single-command demo, output screenshot/GIF.
- Framework — opinionated structure others build on (Next, Rails, etc.). Weight: "why this over alternatives," opinionated example, conceptual model.
- App / service — runnable thing, not imported (self-hosted, web app). Weight: deploy story, config surface, screenshot.
- Standard / spec / docs — knowledge artifact, not code. Weight: scope statement, navigation, change log.
- Personal project / experiment — explicitly unstable. Weight: honest scope, "what this isn't," low-friction trial.
State the inferred category back to the user with one-line evidence ("inferred CLI tool — bin/ entry in package.json, README opens with a $ npx line"). In Auto mode proceed silently; Semi-auto/Manual let the user override.
1.5 Audience inference
Who is this README written for? Infer from jargon density, claimed prerequisites, comparison points named.
- Working developers in the same domain — assumes vocabulary, optimizes for "do I want this in my project."
- Newcomers to the domain — explains terms, shows simple example before deep one.
- Maintainers / power users — assumes deep familiarity, optimizes for reference.
- Mixed — most common; means the README needs explicit on-ramps for each.
A README that's pitched at the wrong audience for its category is the single most common shape problem. Catch it here.
Phase 2 — Audit (the rubric)
Each criterion: score 1–5 with one-line evidence and (for any score ≤3) one-line fix. Default 3, evidence required to move.
2.1 The "what is this" sentence
The first sentence after the title should answer "what is this thing" without prior context. Test: imagine a stranger landing here from a Hacker News link. Do they know what it is in 10 seconds?
- 5: One sentence, no jargon, names the category and the differentiator
- 3: Says what it is but buries the "why" or assumes prior context
- 1: Slogan, brand-voice, or lists features before saying what it is
2.2 Hero presence above the fold
Above-the-fold = first ~25 lines, before any heading deeper than H2. Does the README open with an image, GIF, or visible-output block that makes the project's value legible without reading?
- 5: Visual hero matches the project's identity (CLI: terminal screencast; UI: screenshot; framework: code sample)
- 3: Has a visual but it's generic (logo only, badge row pretending to be a hero)
- 1: No visual; opens with prose or badge sprawl
2.3 Install-to-first-success length
Count the lines (or clicks) from "I want to try this" to "I see the thing working." Less is more. Measure as: line of first install command → line of first runnable example → line of first observable output.
- 5: ≤ 3 lines, all copy-pasteable
- 3: 4–8 lines, requires switching contexts (file edits)
- 1: > 8 lines, or requires a full project setup before any output