Tool: design review with scores (impeccable critique)
What it does
- A design review: does this work for people, and does it feel made for this product?
- It runs two separate assessments, then combines them into a scored report.
- A: a design review.
- B: the rule checker plus evidence from a browser.
When to reach for it
- Something is built and you need to know whether it works for people.
- Before deciding between
polish and a redesign.
- To track whether a page gets better over several rounds.
How to use it
- Run
/impeccable critique <target>. The source's examples: /impeccable critique the pricing cards, /impeccable critique the onboarding flow.
- A and B run as separate subagents when the tool allows. A finishes before the checker's output reaches the summary, because that output steers judgement.
- Without subagents, the report must open with a
⚠️ DEGRADED: single-context (<reason>) banner. A degraded critique without it counts as failed.
- The report runs in this order:
- The verdict: authored for this product, or interchangeable with others in the category?
- Two or three strengths.
- Three to five priority issues, each tagged P0 to P3, with what it is, why it matters, the fix, and which command to use.
- Scores against Nielsen's ten heuristics (rules of thumb for usable design), with bands from Excellent to Critical. See Drawer 6.
- A mental-load check and tests with named users. See Drawer 6.
- Two to four questions with concrete options, last.
- It is direct. "The submit button", not "some elements". No "consider exploring…".
- It saves a snapshot under
.impeccable/critique/ and shows a trend over the last five runs. polish picks up its P0 and P1 issues from there.
- impeccable says skipping subagents without the banner is this command's most common failure (asserted, no evidence shown).
- The source's form is
/impeccable <command> <target>, typed in your agent ($impeccable on Codex). Narrow targets get better results: one section beats a whole page.
What happened when we used it
- We used its anti-pattern question (would someone say "AI made this"?) by hand. It found three tells: a number on every section, a small capitals label over every section, and a cream background.
- Two of those were the look Vince had liked, so they were treated as a taste call. A de-tell version was built next to the original to compare.
Source