Think Like Roy review pass: score work against a standard it must cite

Review passFreshModel-agnosticverified 2026-07-31sha a4b116cd43bb

No re-run note on file — that date records when this was last checked, not what the check showed.

Born in: the post where it earned its place →

Breaks when: It is a first pass, not the verdict – the standard's own rule stands: if you haven't seen it run, it's not done, so this never replaces running the work. It needs the full standard pasted (roughly 50KB); a client that silently truncates long prompts will "review" against half a lens and still sound confident – the repo's review-gate.md is the truncation-safe alternative. Scores drift without evidence: on thin input (a lone screenshot, a partial diff) the model can rate axes it cannot actually see – treat any score it can't tie to visible evidence as NOT RUN – and second-model reviewers over-escalate documented design intent into findings (observed live; filtering defect from disagreement stays your job). The cite-the-heuristic discipline is also only as good as the pasted copy: an edited or stale standard yields confident citations to rules that no longer say that.

This is the review prompt that ships inside Think Like Roy, my QA standard for software that has to feel made by people – free since July 2026, CC BY 4.0, at [github.com/royazr/think-like-roy](https://github.com/royazr/think-like-roy). Grab `think-like-roy.md` from the repo (or `review-gate.md`, the compact extract, where a 50KB paste risks silent truncation), paste it where the last line says, attach the work, and any capable model becomes a first-pass reviewer that has to cite the heuristic behind every finding and score the five ship-gate axes, where any zero blocks the ship. Born from too many “looks fine” reviews by reviewers who never ran anything. (Typographic dashes from the repo copy are normalised to plain hyphens here to pass this library’s byte-lint; the prompts are otherwise identical.)

You are reviewing work against the "Think Like Roy" QA lens (pasted below).

Review this <diff | screenshot | screen>: {{work}}
Context: platform = {{platform}}, surface = {{surface}},
user state = {{user_state}}.

Rules:
- Use ONLY the lens. Every finding must cite the heuristic, principle, or named
  tell it comes from. No taste-only opinions.
- Score the five ship-gate axes (trust, usefulness, human-feel, native-feel,
  accessibility) 0-2 each; any 0 is a BLOCKER.
- Hunt the AI-house-style tells explicitly - in the work AND in your own
  wording. Don't write the slop you're paid to catch.
- Output in the Findings format (The read / Findings by severity / What works /
  Next move). Flag anything you could not verify from what was given - don't
  invent evidence.
- Score only what you can verify from the inputs. An axis you cannot see
  evidence for is NOT RUN - never guess it a 0-2. You are a first pass,
  never the sole approver.

<paste the full think-like-roy.md document here>

Fill in: work platform surface user_state

Shared as is, for reference. Read it and decide what it will do before you run it; using it is your responsibility, under the terms.