In the Skill's Own Terms
Improve UI audits one coherent product surface against the design system that actually governs it, then writes self-contained implementation plans for another agent to execute. It is read-only on product source: files are created or edited only under design-plans, and no dependencies, formatters, commits or pushes are allowed.
Findings must survive three proofs, contract, runtime and correction, plus a later falsification pass, and the report stops at three, ordered by confidence, user impact, reach and correction cost. It suits designers and developers reviewing, refining or cleaning up an interface without replacing its identity, investigating design-system drift, or preparing a design handoff. Accessibility, behaviour, performance and architecture issues are out of scope unless requested.
How It Works
- 01Select One Coherent Surface
A broad request is narrowed to one deployable application and one coherent surface family for a primary product task; the skill states that selection instead of treating the repository as a single product.
- 02Reconstruct the Local System
It uses a source only after proving it is current and governs the selected surface; drafts, proposals, migrations and task lists describe future intent unless explicitly accepted and current.
- 03Require Three Proofs per Candidate
Every kept candidate needs a contract, a runtime path and one deterministic correction; search results, repetition and implementation differences produce candidates, not findings.
- 04Vet and Falsify
Candidates are attacked again before reporting, and only the ones that survive that pass are allowed into the findings table.
- 05Report, Ordered and Capped
Findings are ordered by confidence, user impact, reach and correction cost, and the list is cut at three.
- 06Specify Selected Changes
Before writing, cited sources are re-opened, the current commit is recorded when available, and exact reusable primitives and exemplars are identified across the affected surfaces.