Md Review

Markdown Review Gate

The quality gate that runs before Markdown becomes HTML. A converter will happily render a document with three H1s, four dead links, an image with no alt text, and a 67-word sentence — the HTML validates and the page is still bad. This skill catches those defects while they are still cheap to fix, and fails the build when they are blocking.

Zero network calls, by design. Relative links and anchors resolve on disk; external URLs are inventoried and reported but never fetched. A gate that fails because someone else's server was slow is a gate engineers learn to ignore.

When to use this skill

• Before publishing a doc, guide, or article that will be converted to HTML
• Wiring a docs CI gate that must block a merge on real defects without blocking on style
• Auditing an inherited Markdown corpus to size the accessibility and link-rot backlog
• Enforcing house terminology across a docs set (front-end vs frontend, GitHub vs Github)
• Checking accessibility of source content against the WCAG criteria that survive conversion
• Calibrating prose to an audience — a runbook read at 3am needs a different band than an API reference

Inputs the skill expects

• One or more Markdown files (with or without YAML frontmatter)
• A review config JSON: required frontmatter fields, structure and accessibility thresholds, term map, gate settings
• The target audience for the prose (drives the readability band — the single most consequential input)
• The project root, when the corpus uses root-relative (/docs/...) links
• The blocking policy: which severity fails the build, and the warning budget

Clarify First

Before generating, confirm these inputs. If any is unknown or vague, ASK — do not assume:

• Target audience for the prose — why it changes the output: selects the Flesch band and grade ceiling; a general-public band (60-80) and a specialist band (40-60) flag opposite sets of sentences
• Required frontmatter fields — why it changes the output: every missing field is an error, so guessing the schema produces either false blockers or a silent gap
• Blocking severity and warning budget — why it changes the output: decides whether the run reports or blocks, which determines whether this is an audit or a gate
• Whether the corpus is new or inherited — why it changes the output: an inherited corpus needs report-only phase 1, not a gate that fails on day one

Stop rule: ask only the 2-3 that most change the output. If the user says "just draft it," proceed and list your assumptions at the top of the artifact.

Workflows

Workflow 1 — Gate a document before publication

The default path. Run all three tools; any non-zero exit blocks.

1. Pick or write a config profile (start from assets/sample_review_config.json).
2. Run the structure/frontmatter/accessibility gate.
3. Run the offline link checker.
4. Run the readability and terminology scorer.
5. Fix errors; triage warnings against the budget.

cd "$(git rev-parse --show-toplevel)"
CFG=markdown-html/md-review/assets/sample_review_config.json
DOC=markdown-html/md-review/assets/sample_article.md

python3 markdown-html/md-review/scripts/md_review_gate.py --input "$DOC" --config "$CFG" --format text
python3 markdown-html/md-review/scripts/link_checker.py --input "$DOC" --root "$PWD" --format text
python3 markdown-html/md-review/scripts/readability_scorer.py --input "$DOC" --config "$CFG" --format text

The shipped sample_article.md deliberately contains real defects, so this run exits non-zero. Swap in sample_article_clean.md to see all three pass.

Workflow 2 — Audit a corpus without blocking anything

Phase 1 of any rollout. Collect the real finding distribution before deciding what to enforce.

1. Run every file with --fail-on never so nothing exits non-zero.
2. Emit JSON and append to a single JSONL stream.
3. Rank rules by frequency; tune thresholds and the term map before switching the gate on.

cd "$(git rev-parse --show-toplevel)"
CFG=markdown-html/md-review/assets/sample_review_config.json

find markdown-html/md-review/assets -name '*.md' -print0 |
  xargs -0 -I{} python3 markdown-html/md-review/scripts/md_review_gate.py \
    --input {} --config "$CFG" --format json --fail-on never > /tmp/md_audit.jsonl

python3 -c "import json;[print(f['rule']) for l in open('/tmp/md_audit.jsonl') if l.strip().startswith('{')]" 2>/dev/null || \
  echo "inspect /tmp/md_audit.jsonl for the per-file finding arrays"

Workflow 3 — Calibrate prose to an audience

When the complaint is "nobody reads our docs" rather than "our docs are broken".

1. Score the document and read the sentence-level findings, not just the aggregate.
2. Rewrite the very-long sentences first — they dominate the score and the reader's experience.
3. Re-score and confirm the long-sentence percentage is under 10%.

cd "$(git rev-parse --show-toplevel)"
CFG=markdown-html/md-review/assets/sample_review_config.json

# Full report, including passive-voice and terminology findings
python3 markdown-html/md-review/scripts/readability_scorer.py \
  --input markdown-html/md-review/assets/sample_article.md --config "$CFG" --format text

# Readability band only — ignore terminology while rewriting sentences
python3 markdown-html/md-review/scripts/readability_scorer.py \
  --input markdown-html/md-review/assets/sample_article.md --config "$CFG" \
  --fail-on readability --no-passive --format json

Decision frameworks

Severity assignment — what earns an error [PROVEN]

A finding blocks publication only if it passes all three tests. Everything else is a warning.

Broken link, missing alt text, skipped heading level, headerless table, missing required frontmatter field → error. Long sections, terminology drift, heading capitalization, readability band → warning. Passive voice → info.

Never downgrade a11y.missing-alt. Every other rule has a defensible exception; this one does not. If an image is decorative, mark it decorative — do not suppress the rule.

Readability target bands by audience [PROVEN]

The single table that makes readability scoring useful. A score without a target audience is noise.

Only the floor blocks. Prose easier than its band is prose more people can read; the scorer records it as info. Gating both bounds teaches authors to pad sentences, which inverts the point.

The safety-critical row is the one teams get wrong. Comprehension collapses under stress — an incident runbook written at grade 12 is unreadable at 3am during an outage.

Sentence-length thresholds [PROVEN]

Target mean ≤ 20 words with ≤ 10% of sentences over 30. The percentage matters more than the mean — an 18-word average with 20% monsters reads worse than a 22-word average with none.

WCAG coverage — what source-level checks can and cannot prove [RECOMMENDED]

Run the bottom three against converted HTML. Claiming source-level checks prove WCAG conformance is how teams end up with a compliance badge on an inaccessible site.

Rollout sequence for an existing corpus [PROVEN]

Switching a gate on across a legacy corpus in one step fails every time.

Phase 2 carries the value. Gating only the files a change touches makes the gate immediately useful and never blocking on unrelated debt.

Exit code contract [PROVEN]

Keep 1 and 2 distinct. Collapsing them sends every failure to the wrong person first.

Anti-Patterns

The Network Link Checker In The Merge Gate

Mistake: Wiring an HTTP link checker into the blocking pre-merge gate so every external URL gets fetched on every run. Why it happens: Dead external links are a real problem, and checking them feels like the same job as checking internal ones. The tooling usually offers both behind one flag. Instead: Resolve internal targets on disk in the blocking gate — it is deterministic and finishes in milliseconds. Inventory external URLs and verify them in a separate scheduled, non-blocking job. A gate that intermittently fails on someone else's 503 gets re-run reflexively within two weeks, and then nobody reads the real failures either.

Gating On The Readability Ceiling

Mistake: Failing the build when a document scores above its target Flesch band, on the theory that the band is a specification to hit. Why it happens: The band is written as a range, so both ends look like thresholds. Treating it symmetrically feels rigorous. Instead: Block only on the floor. Prose easier than its audience requires is a win, not a defect — record it as info. Teams that gate both ends get authors padding sentences with subordinate clauses to climb back into range, producing exactly the writing the metric exists to prevent.

The Term Map That Only Grows

Mistake: Adding every style disagreement to the terminology map and never removing anything, until the map has 400 entries and every document produces twenty warnings. Why it happens: Adding an entry is a one-line fix that closes a style argument permanently. Removing one requires re-litigating it. Instead: Cap the map at the terms that actually matter — product names with canonical capitalization, contested hyphenation, deprecated names, inclusive-language replacements — and review it quarterly. Every entry should have a reason someone can state out loud. Keep terminology at warning severity; blocking a release on Github teaches authors the gate is petty, and a gate perceived as petty gets bypassed.

Zero Warnings On Day One

Mistake: Adopting the gate with max_warnings: 0 against an inherited corpus, producing a 400-finding first run. Why it happens: Zero is the obviously correct end state, and starting anywhere else feels like tolerating defects. Instead: Set the budget at the current warning count, then ratchet down 10-20% per quarter. A first run that produces one enormous cleanup PR gets rubber-stamped, not reviewed, and the debt returns within a release. Phase the rollout: observe, then gate changed files, then ratchet.

Inline Suppression Comments

Mistake: Adding <!-- md-review-disable a11y.missing-alt --> markers in documents to silence findings the author disagrees with. Why it happens: It unblocks the immediate merge and feels surgical compared to changing the shared config. Instead: Fix the config, downgrade the severity, or disable the rule globally with a recorded reason. Inline suppressions spread by copy-paste, are never reviewed, and become permanent exemptions nobody can justify. If a rule needs suppression often enough to want an inline escape hatch, the rule itself is wrong — change it once, in the open.

Files