coach
Interpret recurring measured findings and honest trends without turning taste into a skill score.
Audit current renders and stress coverage, repair real defects and stop under the bound refinement policy.
$ npx -y skills add 3x-haust/oh-my-design --skill audit --agent claude-codeHow it fires
How this skill gets triggered: by you, by Claude, or both.
/auditContext preview
The summary Claude sees to decide when to auto-load this skill.
Audit current renders and stress coverage, repair real defects and stop under the bound refinement policy.
name: audit description: Audit current renders and stress coverage, repair real defects and stop under the bound refinement policy.
Read only the current visual-measurement/slop/refinement contract. Measure the actual build, inspect every surface/state/view and separate stress capture, then launch the host's isolated mode-bound Eye. Blind packets exclude references, maker rationale, prior scores and budget pressure. Native signatures establish provenance, not aesthetic truth or human usability.
Apply the route-bound quality floors independently. Required evidence, task/access/safety/material defects and failed stress cannot be budget-waived. Accept an initially clean result without inventing a repair round. Optional polish may plateau under the bound policy; do not relabel known defects as advisory.
Repair through the responsible owner, recapture and compare distinct authenticated outcomes. Regression requires trusted rollback and fresh observation. Three failed post-repair observations of the same stable defect ask the user/needs-human; wording changes, duplicate reads and restart do not reset it. Typed pause stops continuation.
Finalize only through the actual native publisher and guard completion. Report current verification and honest debt/limits; a stopped incomplete run is not released work.
AI that designs like a human. Give OMD one request. It researches, compares real material, chooses a direction, builds, renders, and repairs internally.
Interpret recurring measured findings and honest trends without turning taste into a skill score.
Dispatch the existing candidate-generation stage from minimal tokens to one authentic direction choice.
Review actual current design evidence and rank observable problems without implementing changes.
Expand the selected sample into the complete requested surface, state and viewport scope.
Implement an already chosen Figma target faithfully while verifying unrepresented behavior, states and responsive gaps.