improve-brandlisted

# improve-brand — fan-out DSL improvement loop Driven by a verify-map. For each layout whose `struct_diff_ratio` exceeds a threshold after a verify-loop run, dispatch **one sub-agent per layout** in parallel. Each sub-agent gets the per-slide diff overlay, the current DSL, and a tight instruction set; it edits the DSL only; the parent re-runs the loop and reports plateau-vs-progress. ## When to use - You have a brand pack (`brands/<brand>/` with `tokens.json`, `layouts/*.slide.dsl`, `verify-map.yaml`) and a source PPTX deck to match. - The pack is past the scaffold stage (every layout already builds and renders something recognisable). If it doesn't exist yet, start with `compile-html` — improve-brand is a polishing loop, not a scaffold tool. - You want to drive every layout's `struct_diff_ratio` to ≤5% (or `--threshold`). ## Checklist You MUST create a task for each step and complete them in order. Full detail (commands, dep notes, the process diagram) is in [`references/loop-detail.md`](references/loop-detail.md). 1. Run `brand_verify_loop.py … --snapshot-baseline` (builds/renders/diffs every layout; locks the before-renders). 2. Read `<output-dir>/diff/report.json` for per-layout `struct_diff_ratio` + overlay paths. 3. Filter to the work set — layouts above `--threshold`. 4. Fan out sub-agents — ONE message, one `Agent` call per layout, **in parallel** (prompt template: [`references/per-slide-prompt.md`](references/per-slide-prompt.md)). 5. Wait for every sub-agent;