Runs the selection and revision discipline that separates shipped AI assets from lucky ones - a minimum four-candidate set per final asset, a tell checklist covering rendered-text read-back, hands, logo geometry, shadow, depth-of-field, and perspective agreement, a draft-then-premium spend ladder, and an edit-not-re-roll state machine after base approval. Use when someone asks "how do I know if this AI image is good enough to ship", "the client keeps spotting weird AI artifacts in our images", "our AI generation costs are out of control", "every revision changes things the client already approved", or before any generated asset goes into a final deliverable. Do NOT use to judge creative quality against a bar - use creative-quality-ladder instead; to fix the prompt that produced the flaw, use photoreal-prompt-craft.
---
name: generation-qa
description: Runs the selection and revision discipline that separates shipped AI assets from lucky ones - a minimum four-candidate set per final asset, a tell checklist covering rendered-text read-back, hands, logo geometry, shadow, depth-of-field, and perspective agreement, a draft-then-premium spend ladder, and an edit-not-re-roll state machine after base approval. Use when someone asks "how do I know if this AI image is good enough to ship", "the client keeps spotting weird AI artifacts in our images", "our AI generation costs are out of control", "every revision changes things the client already approved", or before any generated asset goes into a final deliverable. Do NOT use to judge creative quality against a bar - use creative-quality-ladder instead; to fix the prompt that produced the flaw, use photoreal-prompt-craft.
---
# generation-qa
Without a selection gate, every asset that ships is a coin flip the audience gets to grade. Every model leaderboard that matters is pairwise human preference - selection is the craft signal, and sampling variance means the fourth candidate is frequently the dramatic winner at near-zero marginal cost. The tell checklist below catches exactly the artifacts audiences use to detect AI - the six-fingered hand in the launch gallery - while the spend ladder stops the opposite failure: burning premium credits on exploration a draft tier answers for a fraction of the price. This is the gate every generated still and clip passes before it enters a final deliverable.
## Operating procedure
Run the steps in order: budget before generating, drafts before premium, checklist before selection, approval before edits. Each gate exists to make the next one cheap - a priced ladder makes candidates affordable, screened candidates make selection fast, and an approved base makes every later change a targeted edit instead of a fresh gamble.
### Step 1: Gather the QA inputs
Elicit exactly these before screening anything. If any input is a guess, label it a guess and move on.
0. **Charter**: the client's brand charter block, or the campaign consistency kit from brand-consistency-prompting. Paste the charter block if one exists. If the client has a charter but the foundations pack is not installed, elicit the needed values directly and label them provisional. If no charter exists at all, halt and route to premium-design-foundations - never substitute worked-example values.
1. **Manifest**: the routing manifest from image-model-router - asset counts, lanes, which lane generates which deliverable.
2. **Unit costs**: draft-tier and premium-tier unit cost per lane (per image, per second of footage). Label guesses.
3. **Intended copy**: the exact text every asset carries, verbatim - this is the read-back reference.
4. **Geometry references**: the logo and product shots the geometry check compares against.
5. **Optical claims**: the prompt's stated focal band and aperture per asset, from photoreal-prompt-craft - the optics-agreement rows check the output against these claims.
6. **Approval workflow**: who signs the base, because the edit-not-re-roll state machine activates at that signature.