| name | exemplar-review |
| description | Review a diff against the best exemplars for its domain — checking both whether it reinvents a community-standard tool or idiom (prior-art), and whether it clears the quality bar a recognized exemplar would hold (excellence-bar, e.g. Apple-grade UI, Stripe-grade docs). Use when the user wants a taste, quality, or idiom pass, asks "is this <X>-grade?", "what would <person> think?", "are we reinventing the wheel?", "is there a standard way?", or runs /exemplar-review. Complements correctness/simplify/maintainability reviewers with the perspective they structurally lack. |
exemplar-review
The reviewer that asks two questions local reviewers can't see:
- Prior-art — Is there a community-standard tool, library, or idiom a fluent practitioner would reach for, that this diff reinvents or diverges from?
- Excellence-bar — Does this clear the bar a recognized exemplar would hold it to? (Apple-grade UI, Stripe/Vercel-grade docs, Jobs-grade scope, Reitz-grade API ergonomics.)
Correctness/simplify/maintainability reviewers optimize within the chosen approach. This one questions the approach and grades the taste.
Workflow
- Read the diff and classify it by artifact × dimension (filetypes, imports, frameworks, what changed). See REFERENCE.md for the routing table.
- Pick 1–2 exemplars and state, in one line each, why this diff routed to them (cite the signal). Don't sprinkle names — two well-chosen beats five.
- Run Mode A (prior-art) for each: is there an established tool/pattern? Name it, show its canonical usage, say what to replace and what it deletes.
- Run Mode B (excellence-bar) for each: decompose the exemplar's standard into a rubric of observable criteria (see REFERENCE.md templates), score the diff against each, give a specific fix per gap. Never a bare verdict.
- Synthesize: rank the findings by leverage; mark each with a confidence and the tradeoff it carries.
Non-negotiable rules
- Lens, not oracle. Ground every finding in the exemplar's citable, shipped work or documented principle ("requests tests against httpbin, so…"; "Apple HIG requires an empty state"). Never fabricate "X would say." The name selects the idiom set; the technical merit is the proof.
- Mode B emits a rubric, never a vibe. "Make it more Apple" is banned. Decompose into observable criteria → score → concrete fix. Exemplars with public, observable output (Stripe docs, Apple HIG, a maintainer's library) work; a generic "genius" persona drifts to motivational fluff — refuse to route to one.
- Respect constraints. Aspirational ≠ ignoring the repo's conventions, the user's documented style, or stated scope. A fix that violates a hard constraint is not a fix; name the tension instead.
- Guard against authority bias. A famous name does not make a finding true. If you're not confident, say so. Defend drops on merit.
- Read-only when run inside another reviewer. Under
/review-fix, only report; never edit, commit, or push.
Output
Per exemplar:
### <Exemplar> — chosen because <signal in the diff>
Prior-art:
- <standard/tool> — diff does <X>; replace with <canonical usage>; deletes <Y>. [confidence]
Excellence-bar (rubric):
| criterion (observable) | bar | current | fix |
| ... | ... | ... | ... |
Grade: clears / falls short on <n> criteria.
Close with a ranked, deduped Top fixes list (leverage-ordered), each one line: finding, fix, confidence.
Modes of invocation
- Standalone:
/exemplar-review on the current diff (or a named target).
- Embedded in an orchestrator (e.g. as an extra reviewer under
/review-fix): report-only; return the Top fixes list (one numbered finding per line, each with its fix and confidence) so the orchestrator can triage alongside the other reviewers.
Litmus
It's working when it surfaces things the other reviewers can't — both "delete this, use the boring standard tool" and "this empty state isn't Stripe-grade; here's the missing copy and the fix." If a run only restates correctness/style nits, it routed wrong or stayed too generic.
