Start Agent Skill

<!-- HARNESS MIRROR — do not edit here. Canonical source: skills/ or commands/. After editing source, copy into cli/assets/<harness>/ and repo-root harness mirrors. -->

Context: this sub-skill is one lens of the broader ui-craft skill. If the ui-craft skill is also installed, read its SKILL.md first for Discovery + Anti-Slop + Craft Test, then apply the specific lens below.

Orient the user. Load the ui-craft skill, read the current project, and report what ui-craft can do right now for it. This is the front door — most people run it first.

ORCHESTRATE-ONLY RULE

This command detects state and routes. It MUST NOT build, edit, audit, or re-implement any phase logic. It contributes only:

• Project detection (read-only).
• A "what's available now" map across the three usage layers.
• One recommended next step + the command to run for it.

When the user wants to act, hand off to the real command (/brief, /tokens, /sddesign, /craft, /finalize, the review agents, etc.). Do not inline their steps.

Step 1 — Detect (read-only)

Scan the project. Do not write anything. Gather:

| Signal | How to detect | Used for |

|--------|---------------|----------|

| Framework / styling | package.json deps, config files, file extensions — Tailwind, CSS Modules, styled-components/Emotion, vanilla CSS, Vue/Svelte/Astro SFC | Confirm ui-craft adapts to this stack |

| Token spine | CSS vars (--color-*, --accent-*, --font-*), Tailwind theme.extend, a token file | Gate: is /tokens needed? |

| Brief | .ui-craft/brief.md exists? | Gate: is /brief needed? |

| External design contract | DESIGN.md or design-tokens.json at repo root (ecosystem convention some teams use) | Treat as brief + token input — read it, honor it, don't duplicate it |

| Spec | .ui-craft/spec.md exists? which surfaces? | Is there work in flight? |

| Harness | Claude Code (plugin/skill) vs .codex/.cursor/.gemini/.opencode/.agents mirror | Which layers are reachable (agents/MCP are Claude Code / MCP-client only) |

| Existing UI | are there components/pages already, or is this greenfield? | Build-new vs review/polish framing |

Keep detection fast and quiet — a few reads, no deep traversal. Report findings as a short table.

Step 2 — Report the three layers

Print what's available now, mapped to how people actually use ui-craft. Mark each layer ● active / ○ available / – not reachable in this harness.

ui-craft is a design engineering SYSTEM. You're set up for:

● Layer 1 — Just ask
    Describe any UI and the agent designs with taste (anti-slop, system tokens,
    real hierarchy). No command needed. This is the default and the point.

○ Layer 2 — Drive it
    Slash commands for focused passes. Given THIS project:
      <tailored list — see Step 3>

○ Layer 3 — Verify it
    design-reviewer + a11y-auditor agents · ui-craft-mcp gates · UICraftScore · ui-craft-detect CLI.
    <mark – if the harness can't reach agents/MCP, and say so>

Tailor the bullets to the detected state — don't print the generic catalog.

Step 3 — Recommend ONE next step

Pick the single highest-value next action from detected state. Use this decision order:

• No brief, non-trivial project → recommend /brief ("anchor every later decision to your product's principles — 5 min, writes .ui-craft/brief.md").
• Brief exists, no token spine → recommend /tokens ("establish the 3-layer token spine so builds match a system, not inline guesses").
• Brief + tokens exist, building something net-new → recommend /sddesign <surface> (full pipeline) or /craft <surface> (one-shot) — ask which.
• Existing UI to improve → route by intent: review → ui-craft:design-reviewer + ui-craft:a11y-auditor (Claude Code) or /critique + /audit; polish → /polish; simplify → /distill; flat/generic → /bolder; loud/busy → /quieter; dated site to modernize → /redesign.
• About to ship / merge → recommend /finalize (10-pass gate) and, for CI, npx ui-craft-detect or the score_ui MCP tool.

Print it as one clear line:

> Recommended next: … — <one-sentence why>.

Then offer the full menu only if asked: "Want the full command list, or shall I run that now?"

Step 4 — Hand off

If the user confirms, run the recommended command. Otherwise stop — /start never acts on its own. Its job is done once the user knows where they are and where to go.

Notes

• Read-only. /start writes nothing and changes no code. The artifacts (brief.md, spec.md, tokens) are written by their own commands.
• Harness honesty. If the project is in a non-Claude-Code harness (.codex, .cursor, etc.), Layer 3 agents and the MCP server may not be reachable from that agent — say so plainly and point to the CLI (ui-craft-detect) as the portable verify path.
• Craft-intent routing. Full-surface builds (/craft, /sddesign) declare a Craft Read before code. If the user describes a built UI as "generic", "template-y", or "too safe", route to /bolder; if "too loud" or "too busy", route to /quieter.
• Don't re-run Discovery. If a brief or spec already exists, don't interrogate the user for preferences — read the artifacts and report.

Next step: Follow the rung this run named. /start is read-only — the value is in doing the one command it picked.