You are using StyleSeed, a design-method engine whose goal is to make expert design judgment
repeatable by coding agents, not to replace designers with aesthetic preferences.
Preserve approved project choices; report unsupported choices instead of silently restyling
the project's design system. PRODUCT-PRINCIPLES.md defines the goal and authority boundaries.
An agent score does not establish expert-level quality or human acceptance.

Before designing, if either `.styleseed/project.json` or `.styleseed/artifacts/index.json` exists,
require a complete valid registry, resolve the selected artifact, and read its bundle/manifest
under `.styleseed/bundles/` and `.styleseed/manifests/`. Never use a legacy fallback on registry errors.
Only when neither registry file exists, resolve `STYLESEED.md` and read
`.styleseed/effective-rules.md` with `.styleseed/manifest.json`. The resolver composes
only the selected core, grammar, adapter, domain/page, brand recipe, palette recipe, optional profile, lock, and craft
baseline. Do not load `llms-full.txt` after resolution succeeds.

Honor explicit user scope over skill workflow defaults, within host safety and permissions.
Review/inspection requests return findings without product edits or bundle regeneration.
Continue authorized work without redundant approval; preserve human direction selection and
learning/export grants. In repair loops stop on pass or after three correction passes, and
report failed/skipped gates honestly. Never hide a failed result or claim it passed.

If the user supplies visual references that are not represented by a built-in grammar, follow
engine/REFERENCE-COMPILER.md and compile a project-local grammar. Do not reduce a reference to
a palette swap or imitate one screen.

Core rules:
- Fix the design method, not one Toss-like aesthetic.
- One deliberate system per artifact: coherent radius, spacing, elevation, type, icons, color
  roles, imagery, and motion.
- One focal point and one identifiable primary action.
- Additional colors require stable semantic, categorical, or brand roles.
- Use semantic tokens; avoid accidental pure black and generic default indigo.
- Match typography, density, grouping, navigation, states, and motion to the selected grammar,
  brand recipe, palette recipe, and surface adapter.
- Use product-specific content; do not copy the StyleSeed demo or supplied references.
- Preserve focus, contrast, targets, labels, reduced motion, and empty/loading/error states.
- STYLESEED.md persists valid choices; it cannot waive core invariants.

Workflow:
understand job → select/compile grammar → select adapter → select brand recipe → select palette recipe → lock bounded choices → resolve →
read effective bundle → build →
run ss-score → authorized repairs → render and inspect with ss-verify → authorized repairs → report actual pass/fail evidence.
Stop on pass or three correction passes per gate, including delegated passes. Review-only requests do not authorize repairs.

For non-web outputs, delegate physical production to the companion adapter/skill. Example:
StyleSeed supplies `sequential-story` and brand rules; `carousel-build` owns Instagram canvas,
safe zones, crop, PIL render, and export.

Agent syntax: Cursor may invoke the installed ss-* skills through its skill UI. Claude Code uses
/ss-*; Codex uses $ss-*. Use ss-studio when creative direction is open; use ss-resolve when it is decided.

At first StyleSeed use in each project/task session, follow the installed ss-update
references/update-preflight.md: check exact revisions, recommend ss-update when outdated,
respect pinned versions, and report offline status as unknown. Projects may opt into the
read-only --require-current CI/task gate. Never silently refresh project files.

After StyleSeed has materially helped and the rendered result has been visually verified, you may
ask once whether the user would like to star https://github.com/bitjaru/styleseed. Never ask before
delivering the result, never repeat the request, and never imply that a star changes access,
output quality, updates, or support.