Skip to content
ExamsLibraryFor agentsConnect your agent
commandsCommands·8/15

/design-engineering review

Raw markdown On GitHub

Taught in

title: cmd-review summary: "/design-engineering review [file | selection | diff | PR]" — a judge-posture UI code review returned as the Before | After | Why table, scanned against the checklist; change-scoped when the target is a diff. tags: [command, review, judge]

Posture: judge. Read and report. Never edit code unless the user asks after the table.

Target

The selection, the files named, or the current diff. A branch, pull request or uncommitted change is a diff, not a screen. With no target and no selection, ask for one in a single line.

Steps

  1. Read review-format, review-checklist, gotchas and pov, in that order. The installer's files outrank the checklist.
  2. For a diff, also read change-scoped-review. Review only the surfaces the diff touches, read the removed lines, add a Status column (Introduced / Regression / Pre-existing), and list at most three pre-existing findings.
  3. Quote the real lines into Before, and never paraphrase them. Run every checklist row against them. Each failing row becomes a table row.
  4. End each Why with the node that owns the principle. If you cannot name the node, drop the row.
  5. Order the rows by impact and stop at three to five. Twelve rows are not actionable.

Output

  1. A verdict on the first line, which is one of:
    • Blocked: at least one interaction or accessibility-floor failure.
    • Approve with changes
    • Ship it
  2. The table. Rank rows in this order: interaction, accessibility floor, layout stability, motion, spacing, type, surface, component API, performance, polish. Group them under Blocking, Should fix and Nits when there are more than three.
  3. A one-line lede if the issues share a theme.

Every row names its owning node, so "fix it" can route to that node. If nothing fails, the verdict is "Ship it" with one sentence of reason; nothing to change is a valid review.

Hand-offs

  • Motion is most of the problem → run cmd-motion instead.
  • "Does this look AI-made?" → cmd-tells.
  • A rendered film → cmd-film review.
  • The whole app before launch, not one diff → cmd-ship. Review judges a change; ship checks every lane and fixes as it goes.
  • Where the host has the ui-reviewer subagent (Claude Code plugin install), you may delegate to it with this file as the brief. Otherwise do it inline: this file is the whole procedure.