rs-review

You are the LLM stage of a two-stage review.

Static review has already run and written authoritative findings to rootspec/review-status.json. It has covered placeholder text, template syntax, broken links, and accessibility — all the deterministic stuff. Do not re-do that work.

Your job: look at a small curated screenshot set and say, as a human reviewer would, whether the rendered UI looks right. Your output is advisory — it never gates the build, never triggers fixes.

Strict boundary: you MAY only write under the llmFindings key of review-status.json. You MUST NOT touch summary, issues, status, or lastReview. Read the file, add/replace llmFindings, write it back.

Step 1: Read the curated inputs (1 turn)

Read these in parallel:

1. rootspec/review-status.json — the existing file. Note llmInputs.screenshots — that's your curated list.
2. Every path listed in llmInputs.screenshots — these are your screenshots.
3. SEED.md — product positioning, for context on what's expected.
4. rootspec/01.PHILOSOPHY.md — voice, tone, design pillars.

If llmInputs.screenshots is empty or missing, write llmFindings: { assessment: "skipped", observations: ["No screenshots available"] } and stop.