review-templatelisted
Install: claude install-skill DemchaAV/graphcompose-ai-flow
# Review a GraphCompose render
Measure first, then interpret. This skill produces a verdict and a
ranked list of mismatches — it does **not** open a revision, edit the
template, or approve anything.
## When this applies
- The user asks what is still different, or how close the render is.
- A create or revise pass has just rendered and the loop needs a verdict
to decide whether to iterate.
If the user wants the difference *fixed*, this skill produces the
verdict and `revise-template` acts on it.
## Inputs
From the revision folder: `output.png` (and `output-page-N.png`), the
project's `reference/reference.png`, the parent revision's output when
one exists, `layout-snapshot.json`, and the scope recorded in
`revision.json`.
The scope decides what "different" is even measured against — see
[scope routing](../references/scope-routing.md). Reviewing a
`refactor-only` revision against the reference image instead of its
parent answers the wrong question entirely.
Reviewing needs one skill file, not a pack: `visual-regression.md`, per
the pack's `00-loading-map.md`. Reach for a topic file only when a
mismatch turns out to be about that topic.
## Steps
**1. Measure.** Run the deterministic comparison before looking at
anything:
```bash
node scripts/render-and-diff.mjs --project <id> --revision <revision-id> --skip-render [--against parent]
```
`--skip-render` reuses the existing render, so this is the measure step
alone: the reference is scaled to the render's size automat