Ss Build Agent Skill

Build with the composed StyleSeed method

The build method is the product. Score and screenshots are auxiliary evidence, not the source

of design judgment.

Step 1 — Establish the rule set before code

If STYLESEED.md does not exist, run /ss-setup and write it before UI code. If the user

supplied a visual reference that the selected built-in grammar does not capture, run

/ss-reference first. Never reduce an unfamiliar reference to a palette swap.

Then invoke /ss-resolve (Claude Code) or $ss-resolve (Codex), or run its bundled

scripts/resolve-context.mjs --project-root . --from-lock STYLESEED.md --agent <agent>.

Read .styleseed/effective-rules.md and preserve .styleseed/manifest.json as the provenance

record. Do not load llms-full.txt after resolution succeeds.

Step 2 — Compose, do not improvise

The compiled bundle already composes the authority order: core invariants → selected grammar →

surface adapter → domain/page → brand recipe → optional profile → bounded lock → craft baseline. Use the

manifest selection and source hashes to detect drift. Open a full source document only when the

bundle points to a genuine ambiguity; do not reassemble the handbook ad hoc.

Before code, state the effective rule set in one line, for example:

operations-console × SaaS × dashboard × enterprise-workbench × swiss × locked brand tokens

Resolve conflicts by authority. A profile or lock cannot waive task fitness, coherence, or

accessibility.

Step 3 — Build with design judgment

• Make the grammar's user job and primary decision visible in the first viewport.
• Establish one focal point; avoid equal-weight template grids.
• Use the grammar's composition, density, type, color, surface, imagery/data, action, state,

responsive, and motion contracts.

• Use the recipe's containment, geometry, control, collection, and navigation morphology.

Set data-styleseed-recipe="<id>" on the artifact root when using bundled pattern components.

• Use product-specific content and evidence. Never copy the StyleSeed demo or a reference screen.
• Implement loading, empty, error, focus, reduced-motion, and responsive behavior where relevant.

Step 4 — Code gate loop

Run /ss-score on the actual implementation. The score must name the effective rule set and

check both core invariants and grammar-specific tells. Fix the highest-gain failures and

re-score, up to roughly three passes, until ≥80. If it cannot pass, report the real blocker.

Step 5 — Pixel gate loop

For every renderable artifact, invoke /ss-verify: use the adapter renderer, inspect every

required viewport/frame/page and relevant state, fix perceptual failures, and re-render. If no renderer

is available, say the visual gate was skipped; never imply it passed.

Step 6 — Present with proof

Report:

• effective rule set and why it fits;
• final code score;
• visual verification status and viewport;
• material fixes made by the gates;
• STYLESEED.md and any compiled grammar path.

Rules

• Grammar before code; code gate after build; pixel gate last.
• Output grammar is functional. Brand recipe is morphological. Aesthetic profile is optional.

Neither substitutes for the grammar or permits a brand clone.

• The primary action must remain identifiable; additional color is permitted only where the

grammar gives it stable semantic or categorical meaning.

• Re-read the lock and grammar on every UI change.