Ai Content Qc

---
name: ai-content-qc
description: The human quality gate before AI content goes live — the checks a person runs on agent output, the stop conditions, and the publish-first-perfect-later judgement call. Use when asked to QC or review AI content before publishing, set up an approval gate, catch hallucinated statistics, decide if AI output is safe to ship, stop AI slop reaching clients, or someone says 'can we trust what the agent wrote'. For behaviour testing before launch, see eval-writer. For the voice check, see voice-file-builder.
metadata:
  version: 1.0.0
---

# AI content QC

**The human QA gate is not optional.** The agent is fast; you are the one
accountable for what is published under your name. This skill is the gate — the
specific checks, the stop conditions, and the judgement rule that keeps the gate
from becoming a bottleneck.

That judgement rule, from the course: **publish first, perfect later** — a good
post live beats a perfect post in drafts. *Except* for the absolute stops, which
never bend.

## Before you start

| Input | Why |
| --- | --- |
| The output, plus what the agent was given | You cannot check claims against sources you don't have |
| The claims register — which numbers were supplied | Anything outside it is presumed invented |
| The voice file | "Sounds right" is checked against rules, not vibes |
| Where this ships and under whose name | The named person runs the gate, or delegates it explicitly |

## Step 1 — Run the absolute stops first

The course's exceptions to publish-first — these block, always:

1. **Factual errors** — any claim that is checkably wrong.
2. **Invented statistics** — every number traces to the supplied register or it
   goes. "Our CPL dropped" is noise; "dropped 23%" is evidence *only if 23% was
   supplied*.
3. **Plagiarised passages** — nothing republished, only rebuilt from scratch in
   your voice. Not a risk you take.
4. **Broken attribution** — wrong author, wrong byline, missing disclosure.

One of these found → **stop the pipeline and report.** A bad post live is worse
than a late post.

## Step 2 — Run the quality pass

The course's pre-live checklist, generalised:

- **Stats real?** Every number against the register.
- **Links point at things that exist?** Internal links to real URLs — never
  invented ones.
- **Author correct?** The named person, not a default.
- **AI tells gone?** Em dashes (the most recognisable tell), and the voice
  file's banned list.
- **Voice passes?** Checked against voice-rules, not against taste.
- **Claims match what the destination can keep** — the ad's promise, the page's
  delivery.

## Step 3 — Apply publish-first-perfect-later to everything else

If the piece is **90% right and none of step 1 applies — publish.** A live post
can be edited in two minutes. Second-guessing one section for a day costs more
than editing it live. The gate exists to catch the absolute stops, not to
relitigate word choices.

## Step 4 — Feed findings back

Every gate catch is a system fix waiting to happen:

| Caught | Feed back to |
| --- | --- |
| Invented number | The skill's claims rule; the register's visibility |
| Voice drift | voice-file-builder — the tell becomes a rule |
| Same error repeatedly | The skill's steps (skill-writer) or CLAUDE.md rule |
| Generic slop | The writing skill's inputs — usually a persona or voice gap |

A gate that only catches is a cost; a gate that feeds back is how the agent
starts smarter next run.

## Output

```
# QC: <piece> · for: <named person> · destination: <where>

Absolute stops: facts ✓ · stats vs register ✓ · plagiarism ✓ · attribution ✓
Quality pass:   links ✓ · author ✓ · AI tells ✓ · voice vs rules ✓
Verdict: SHIP / SHIP WITH EDITS (live) / STOP — <which absolute, one line>
Fed back: <finding → destination skill/file>
```

## When it breaks

| What you see | What it means | The fix |
| --- | --- | --- |
| Gate takes longer than the writing | Relitigating taste, not checking stops | Step 3: 90% + no absolutes = ship |
| Invented stat reached a client | Register missing or unchecked | No register, no ship; numbers trace or die |
| Same correction every week | Catch never fed back | Step 4: the catch becomes a rule |
| Everything waits days for one approver | Gate is a person, not a process | Named delegate with the same checklist |
| "It sounds off" arguments | Voice checked by taste | Against the voice file's rules only |
| Perfect drafts, empty publish log | Publish-first forgotten | A good post live beats a perfect post in drafts |
| Slop shipped because deadline | Stops treated as negotiable | The four absolutes never bend — late beats bad |

Never let the gate's speed pressure erase the distinction between the two
classes: taste is negotiable and ships at 90%; the four absolute stops are not
negotiable at any deadline.

## Rules

- **The human gate is not optional**, because the agent is fast and you are the
  one accountable for what carries your name.
- **The four absolute stops never bend**, because a bad post live is worse than
  a late post — every time.
- **Numbers trace to the register or die**, because an invented statistic is the
  single most damaging thing an agent ships.
- **Ship at 90% otherwise**, because a live post edits in two minutes and a
  perfect draft earns nothing.
- **Every catch feeds back**, because a gate that only filters is overhead, and
  a gate that teaches is compounding.

## Related skills

- **eval-writer** — tests behaviour before launch; this gates output after.
- **voice-file-builder** — the rules the voice check runs against.
- **copy-compliance-scrub** — the policy-specific gate for paid media.
- **claude-md-writer** — where repeated catches become permanent rules.