quality-check

Benchmark-relative quality gate. Scores the draft against the research dossier's beat spec (depth, consensus coverage, evidence) plus AI-tell and voice signals, runs an adversarial read armed with the

SKILL.md

GitHub

Quality Check Skill

Reads the draft and produces a quality scorecard + a punch list of specific fixes. The question this skill answers is not "does the draft comply with our rules" — it's "a reader opens this article and the #1 ranking result side by side: which one do they keep?" A draft can be perfectly compliant and still lose; that's a FAIL.

History (why this skill looks the way it does): the previous rubric weighted forbidden phrases + voice metrics + BLUF heuristics and let a 1,100-word, 4-item, no-table listicle score 95 against a SERP of 2,500-word, 9-item, table-bearing competitors. Benchmark-relative dimensions now dominate the score.

Input

For slug {slug}:

  • content-pipeline/5-drafts/{slug}.md (the draft to score)
  • content-pipeline/1-research/{slug}.md (the BEAT SPEC + benchmark — required; if it lacks a BEAT SPEC section, flag the dossier as legacy and recommend re-running /research)
  • content-pipeline/3-outlines/{slug}.md (coverage map, word targets)
  • brand-config.md (forbidden phrases, audience)
  • examples/voice/*.md (voice baseline)
  • .claude/skills/draft/references/voice-guide.md

Process

  1. Run the mechanical metrics script:

    python .claude/skills/quality-check/scripts/quality_check.py "<slug>"
    

    (Use --stage cited when re-running after /verify-claims.) This writes content-pipeline/quality-checks/{slug}-metrics.md with subscores for: depth vs benchmark (25), consensus coverage (20), AI tells (25), evidence (15), structure (15) — plus CRITICAL findings and a mechanical score.

  2. Voice + judgment read (your 0–100, worth 40% of the final). Read 1–2 articles from examples/voice/ and then the draft, as an editor. Judge:

    • Would the byline survive on a serious blog? Does it sound like the examples or like an AI?
    • Specificity: does every section teach something concrete, or does it gesture?
    • Product mentions: demonstrated naturally, or bolted on?
    • Information gain: is the [GAIN] section genuinely not on page 1 (check against the dossier's top-page summaries)?
  3. Run the adversarial read — armed with the benchmark. Spawn a Task sub-agent:

    Read content-pipeline/1-research/{slug}.md (note the BEAT SPEC and the top-page summaries), then read the draft at content-pipeline/5-drafts/{slug}.md as a skeptical industry expert who has read every page-1 result for "{keyword}". The brand is {brand}; the audience is {audience}. Answer first: if this draft and the current #1 were side by side, which would you keep, and why — in 3 sentences. Then list the 5 weakest things about the draft relative to what's ranking. Be specific — sentences, sections, missing material. Include 1 thing that genuinely works. Do NOT rewrite or be polite.

    Save to content-pipeline/quality-checks/{slug}-adversarial.md. Be skeptical of all-praise critiques shorter than 200 words — re-run with a sharper brief.

  4. Combine into the report at content-pipeline/quality-checks/{slug}.md:

    • Verdict at top: final score = 0.6 × mechanical + 0.4 × judgment. PASS (≥85 and no CRITICAL finding and no mechanical dimension below 60% of its weight and the adversarial read doesn't conclude "keep the competitor"), BORDERLINE (70–84 or adversarial-negative), FAIL (<70 or any CRITICAL).
    • Metrics summary (numbers, not raw dumps)
    • Adversarial critique
    • Punch list — specific fixes ordered by severity, each pointing at a section
    • Recommendation — proceed to /verify-claims, send back to /draft (voice/prose problems), or send back to /outline / /research (depth/coverage problems — do NOT ask /draft to fix a structural deficit)
  5. On FAIL:

    • Autonomous mode (BLOG_AGENT_AUTONOMOUS=1): don't stop; emit verdict + punch list and return cleanly. The orchestrator owns the retry budget (BLOG_AGENT_REVISION_BUDGET). Route matters: depth/coverage CRITICALs → the revision brief targets the outline and re-drafts affected sections; prose CRITICALs → the revision brief targets the draft.
    • Interactive mode: stop and tell the user what failed and where to re-enter the pipeline.

Output

  • content-pipeline/quality-checks/{slug}.md (combined report, verdict at top)
  • content-pipeline/quality-checks/{slug}-metrics.md (mechanical)
  • content-pipeline/quality-checks/{slug}-adversarial.md (adversarial)

Scoring dimensions

Dimension Weight What kills it
Depth vs benchmark 25 (mech) <70% of target words; item shortfall; missing required table
Consensus coverage 20 (mech) any must-cover topic absent
AI tells 25 (mech) forbidden phrases; crutch phrase ≥4×; uniform rhythm; throat-clearing openers
Evidence 15 (mech) low claim density; <8 real links; naked [link] post-citation
Structure 15 (mech) flat H2 structure; missing intro/conclusion
Judgment overlay 40% of final sounds AI; no information gain; salesy mentions; loses the side-by-side

A perfect-compliance thin article cannot pass. Depth floors bind no matter how clean the prose is.

Why this skill exists

Without it, thinness and AI tells get caught at the board's review — the most expensive place. Catching them here means the orchestrator can regenerate before /verify-claims spends citation work on prose that was going to be rewritten anyway.