bcbeidel/check-readme

Check README

Audit a project's top-level README.md for structural soundness, safety, prose quality, and freshness. The rubric — what makes a README load-bearing in the top 30 seconds — lives in readme-best-practices.md.

This skill follows the check-skill pattern. Tier-1 detection is in 7 scripts emitting JSON envelopes via _common.py (28 rule_ids total). Tier-2 has 7 judgment dimensions read inline by the primary agent. Tier-3 is collision (cross-README duplication, judgment-driven).

When to use

Also fires when the user phrases the request as:

• "check this README"
• "review my README"
• "is this README any good"
• "what's wrong with the README"

Workflow

1. Scope

Read $ARGUMENTS:

• Single path to a README.md — audit that file.
• Directory path — find the top-level README.md (case-insensitive: README.md, Readme.md).
• Empty — refuse and explain.

Refuse on sub-package READMEs (under packages/*/README.md) — different rubric. Confirm scope aloud.

2. Tier-1 Deterministic Checks

Invoke 7 detection scripts:

SCRIPTS="${SKILL_DIR}/scripts"
TARGETS="$ARGUMENTS"

python3 "$SCRIPTS/check_secrets.py"      $TARGETS   # 1 rule:  secret (FAIL)
python3 "$SCRIPTS/check_structure.py"    $TARGETS   # 7 rules: h1-present (FAIL), heading-hierarchy, section-coverage, section-order, toc-threshold, size, prose-line-length
python3 "$SCRIPTS/check_codeblocks.py"   $TARGETS   # 4 rules: fence-language, shell-prompt, smart-quotes, code-line-length (all WARN)
python3 "$SCRIPTS/check_links.py"        $TARGETS   # 3 rules: broken-relative (FAIL), broken-anchor, broken-external
python3 "$SCRIPTS/check_images.py"       $TARGETS   # 3 rules: alt-text, image-size, badge-overload (all WARN)
python3 "$SCRIPTS/check_safety.py"       $TARGETS   # 5 rules: destructive (FAIL), pipe-to-shell (FAIL), tls-disable (FAIL), non-reserved-hosts (FAIL), emoji-headings
python3 "$SCRIPTS/check_completeness.py" $TARGETS   # 5 rules: license-file (FAIL), license-link, contributing-link, todo-markers, readme-gitignored

Each script emits a JSON array of envelopes. recommended_changes is canonical — copy through verbatim.

Script-to-rules map (28 Tier-1 rule_ids):

| Script | rule_ids | Severity | |---|---|---| | check_secrets.py | secret | fail | | check_structure.py | h1-present | fail | | check_structure.py | heading-hierarchy, section-coverage, section-order, toc-threshold, size, prose-line-length | warn | | check_codeblocks.py | fence-language, shell-prompt, smart-quotes, code-line-length | warn | | check_links.py | broken-relative | fail | | check_links.py | broken-anchor, broken-external | warn | | check_images.py | alt-text, image-size, badge-overload | warn | | check_safety.py | destructive, pipe-to-shell, tls-disable, non-reserved-hosts | fail | | check_safety.py | emoji-headings | warn | | check_completeness.py | license-file | fail | | check_completeness.py | license-link, contributing-link, todo-markers, readme-gitignored | warn |

Tier-2 exclusion list. Any FAIL in secret, h1-present, broken-relative, destructive, pipe-to-shell, tls-disable, non-reserved-hosts, or license-file excludes the README from Tier-2.

Missing-tool degradation. check_links.py requires lychee or markdown-link-check for the broken-external rule; when neither is installed, that envelope emits overall_status: inapplicable (other rule_ids continue).

3. Tier-2 Judgment Dimensions

For each README that passed the Tier-2 exclusion gate, evaluate against the 7 judgment rules at references/check-*.md:

| File | Dimension | Severity | |---|---|---| | check-opening-clarity.md | D1 — H1 + one-sentence problem statement orient in 5 seconds | warn | | check-installation-correctness.md | D2 — install steps are copy-pasteable and complete | warn | | check-quickstart-effectiveness.md | D3 — quickstart produces a runnable artifact in <10 minutes | warn | | check-placeholder-discipline.md | D4 — every <placeholder> is justified, not copy-paste padding | warn | | check-warning-prominence.md | D5 — destructive ops carry visible warnings | warn | | check-maintenance-posture.md | D6 — README freshness signals (last-updated, contributor links) | warn | | check-style-and-voice.md | D7 — direct, plain English; no marketing prose | warn |

Evaluator policy: see check-skill-pattern.md §Evaluator policy. Read all 7 rule files first, then evaluate the README in one LLM call.

4. Tier-3 Cross-README Collision

Evaluate against check-collision.md. For multi-README scope (org-wide audits, monorepos), surface duplicate Installation / Contributing / License prose across sibling READMEs as warn. Single-README scope returns inapplicable.

5. Report

Merge Tier-1 (script JSON) + Tier-2 (judgment) + Tier-3 (collision) findings into a unified table:

| Tier | rule_id | Location | Status | Reasoning |
|------|---------|----------|--------|-----------|

Sort: fail before warn before inapplicable; Tier-1 before Tier-2 before Tier-3 within severity. Each Recommendation: line copies through recommended_changes verbatim.

6. Opt-In Repair Loop

Ask exactly once:

"Apply fixes? Enter y (all), n (skip), or comma-separated numbers."

For each selected finding:

• Direct edit — H1 insertion, missing-section addition, code fence language tag, anchor correction. Show diff; write on confirmation.
• Routed to another skill — substantial rewrites → /build:build-readme.
• Tier-2/3 judgment — ask the user; rewrite the section; show diff; write on confirmation.

After each applied fix, re-run the relevant Tier-1 script. Terminate when the user enters n or exhausts findings.

Anti-Pattern Guards

1. Per-dimension LLM call. Collapse into one locked-rubric call per README.
2. LLM-evaluating format compliance. Frontmatter shape, link resolution, line length — handle deterministically in Tier-1.
3. Ambiguous compliance reported as PASS. Surface as WARN (default-closed).
4. Vague finding text. Cite the specific README and exact phrasing.
5. Bulk-applying fixes. Per-finding confirmation required.
6. Re-evaluating scripted rules in Tier-2. Scripts are authoritative for the 28 Tier-1 rules.
7. Suppressing the inapplicable envelope. When lychee/markdown-link-check is absent, the broken-external envelope emits inapplicable — surface it.
8. Embellishing scripts' recommended_changes. Each rule's recipe constant is canonical guidance from readme-best-practices.md.

Key Instructions

• Run Tier-1 deterministic checks first; gate LLM evaluation on structural validity.
• Present all 7 Tier-2 dimensions as a single locked-rubric call per README.
• Include the full README verbatim in every LLM evaluation.
• Tier-3 collision only fires on multi-README scope; single-README scope returns inapplicable.
• Recovery: read-only outside the Repair Loop.

Handoff

Chainable to: /build:build-readme (rebuild non-compliant README from scratch).