Content Quality Check

The QA gate a draft passes before it publishes. It exists because of one fact: readers now pattern-match AI tells in seconds, and every tell that ships spends the brand's credibility. The check exists to catch what the writer can no longer see in their own draft.

It reports before it rewrites. The deliverable is a severity-tagged review with a fix per flag; wholesale rewriting happens only when asked, because the writer owns the draft.

What this consumes

The division of labor across the repo holds: brand-voice-guide builds the standard, this skill enforces it on drafts, and outbound-engine carries its own sequence-specific gate (this skill can still review sequence copy as prose).

The workflow

Four phases. Mechanical first, judgment second, so the model's attention goes where the machine cannot.

Phase 1: Intake and calibration

Read the draft, note channel and audience, and locate the voice directory if one exists (pillars from voice-core.md, banned words from the Avoid section of writing-standard.md). State the bar being applied: universal checklist only, or universal plus brand standard.

Phase 2: Mechanical sweep

Run scripts/check_content.py on the draft, with --avoid pointing at the brand's Avoid words when available. The linter catches everything a machine can catch deterministically: slop phrases, LLM-overrepresented words, "not just X, it's Y" reframes, "let's dive in" openers, announcement cliches, em-dash chaining, rhetorical-question openers, empty-intensifier density, weak-closer formulas, and the brand's own banned words. Same draft, same flags, every run.

Phase 3: Editorial judgment pass

Read the draft as an editor, against references/detection-checklist.md section C, for what no regex can see:

• The opener: does the first line earn attention, or does it throat-clear and stage-set?
• The closer: does it end flat and forward, or restate the opening as a summary?
• Proof: does every claim carry a number, a name, or evidence, or does it float? Flag unproven claims as "needs proof"; verifying truth is the writer's job, naming the gap is ours.
• Specificity: concrete nouns and real examples over abstractions ("the 300-row questionnaire" beats "operational inefficiencies").
• Template shape: does the structure read as a filled-in listicle or a piece with an argument?
• Voice conformance (when a directory exists): would this pass the brand's distinctiveness test, and does each section sound like the pillars or merely avoid the banned words?

Phase 4: Report, then revise on request

Deliver the review: verdict (publish / revise / rework), the linter output verbatim, judgment findings with line references, and a concrete fix per flag. If the user asks for the revision, apply the fixes, re-run the linter until clean, and hand back a before/after so the changes are inspectable. The final gate is mechanical PASS plus an editor's clear.

Scripts

Failure modes to avoid (the quality bar)

• Silent rewriting. The review comes first; the writer decides what to accept.
• Flag-counting without fixes. Every flag ships with its fix, or it is nagging, not QA.
• Enforcing taste as rule. The checklist distinguishes severities; a LOW is advisory and says so. Real lists of three real things are not tricolon abuse.
• Universal-only when a standard exists. If the brand has a voice directory, checking against the generic list alone is a half-done job.
• Missing the structural tells. A draft can pass every word-level check and still read as a template; the judgment pass exists for exactly that.
• Grading truth. "Needs proof" is a finding; fact-checking the proof is out of scope and said so.

Bundled references

• references/detection-checklist.md: the full catalog, severity-tagged: A (mechanical, the linter's spec), B (word and phrase lists), C (judgment checks). Read at Phases 2-3.
• examples/cohere/: a deliberately slop-ridden Cohere announcement draft, the review that catches 20+ flags across both passes, and the revised draft in Cohere's actual voice that the linter passes clean.