The pretty draft is not the finished draft. An AI-assisted paragraph can be grammatical, structured, and still unsafe to publish under your name. It may carry a claim you cannot prove, soften the reader's real problem into fog, or turn a specific offer into polite internet paste.

The Bureau's Anti-Slop workflow treats the draft as material, not evidence. The quality-control pass asks a colder question: what can this page say, what must it prove, and what should the reader do next?

The checklist

Copyable checklist

Run this before publication.

# AI Writing Quality Control Checklist

## 1. Source packet
- [ ] The draft has a reader, use case, allowed facts, proof, exclusions, and missing evidence.
- [ ] Dates, prices, policies, claims, quotations, and tool names have a source or are removed.

## 2. Claim list
- [ ] Every important claim is marked as sourced, experienced, reasoned, speculative, or unsupported.
- [ ] Unsupported claims are sourced, narrowed, cut, or moved to private notes.

## 3. Reader job
- [ ] The opening names the reader's real decision, pressure, or constraint.
- [ ] Each section helps that reader make one clearer move.

## 4. Proof fit
- [ ] The main claim has support that fits its importance.
- [ ] Observed evidence is separated from hypothesis, promise, or preference.

## 5. Voice evidence
- [ ] Voice rules come from samples, not only tone adjectives.
- [ ] The final draft preserves useful specificity instead of sanding it flat.

## 6. Specificity and texture
- [ ] Abstract benefits have been replaced with actions, examples, conditions, or objects.
- [ ] Repeated openers, filler transitions, hollow intensifiers, and grand closers are removed.

## 7. Human critique
- [ ] The sharpest reader objection is named.
- [ ] The smallest necessary repair is made before rewriting the whole piece.

## 8. Final action
- [ ] The reader can see one useful next step.
- [ ] The action matches the proof already shown.
Open the free checklist

How to run the pass

  1. Start with the source packet. Put the facts, proof, missing evidence, exclusions, and reader job beside the draft.
  2. Circle claims before sentences. Check factual, technical, sales, pricing, policy, and performance claims before polishing rhythm.
  3. Repair the main risk first. Do not fix five small phrases while the main promise is still unsupported.
  4. Use samples for voice. Give the model representative passages and refusal patterns so it has evidence, not costume.
  5. Ask for a compact source-status audit. The output should show which claims were kept, narrowed, removed, or left needing proof.
  6. Read the next action last. If the call to action asks for more trust than the page has earned, lower the action or add proof.

A small repair example

Use this rejected example only as a test case. It claims a result the source packet has not supplied:

Before: This workflow helps teams produce high-converting AI content faster and with more confidence.

A source-safe repair names the mechanism instead of borrowing a result:

After: This workflow puts the source packet, claim list, voice rules, and final action in one review path before the draft reaches a client or public page.

The repaired line is less glamorous. It is also easier to inspect. A reviewer can see the objects in the workflow and decide whether the page proves them.

When to use the full Anti-Slop path

Use the deeper path when the draft affects sales, client trust, technical accuracy, publication, or any claim a reader could challenge. The local Anti-Slop kit uses the same order: source integrity first, known pattern gates second, human texture third, then final QC.

For a live draft, start with the free cleanup checklist. To see the repair logic in public, compare the before-and-after examples. If the same problems recur across sales pages, emails, client work, scripts, or offers, inspect the Anti-Slop kit.

Archive note

This field note fits method VB-M003: Productized diagnostic audit because the object is inspectable, the delivery format is fixed, and the value comes from a bounded diagnosis with an evidence trail.