mirror of https://github.com/garrytan/gstack.git
Replace the single "Check for DESIGN.md" step with a priority-ordered capture (live DOM -> code tokens -> DESIGN.md -> sibling apps), treated as the source of truth. A project without a DESIGN.md now still has a real system to calibrate against instead of falling back to inventing one. Close the loop at artifact-generation time: - Never invent a new aesthetic when one already exists: reuse captured tokens verbatim in any mock/redesign/variant; introducing a new theme is itself a HIGH-impact finding. Reinforced inline at the `$D generate --brief` call so it binds at the moment of generation, not only in the preflight. - Design from the shipped app, not the spec, when a real implementation exists. Add two Content & Microcopy checklist items (in the resolver): - Domain vocabulary: flag generic app-speak where the product's own domain has a term; verify against the real schema/spec, never invent. HIGH-impact. - Human register: write like a knowledgeable colleague, not a system log; raw schema/field identifiers leaking into UI copy is the tell. These two are semantic/intent checks, so they live in the design-review methodology, not the code-level review/design-checklist.md subset. Edited source only (SKILL.md.tmpl + scripts/resolvers/design.ts) and regenerated all hosts with `bun run gen:skill-docs --host all`. |
||
|---|---|---|
| .. | ||
| preamble | ||
| browse.ts | ||
| codex-helpers.ts | ||
| composition.ts | ||
| confidence.ts | ||
| constants.ts | ||
| design.ts | ||
| dx.ts | ||
| gbrain.ts | ||
| index.ts | ||
| learnings.ts | ||
| make-pdf.ts | ||
| model-overlay.ts | ||
| preamble.ts | ||
| question-tuning.ts | ||
| redact-doc.ts | ||
| review-army.ts | ||
| review.ts | ||
| sections.ts | ||
| tasks-section.ts | ||
| testing.ts | ||
| types.ts | ||
| utility.ts | ||