/claude-design

Optional hook — deterministic Claude Design hand-off triggering

Skill auto-triggering is probabilistic and under-fires on instruction-shaped hand-off prompts ("Use the claude_design MCP to import this project: …") — measured ~1–2 / 5 even with a tuned description, because the model reads the prompt as a self-contained instruction and decides it can act without consulting a skill. No description wording reliably wins that.

To make the claude-design skill fire deterministically on a Claude Design hand-off, wire detect-handoff.sh as a UserPromptSubmit hook. It reads the submitted prompt (JSON on stdin) and, when it matches a hand-off signature — api.anthropic.com/v1/design/mcp, claude_design mcp, claude.ai/design/p/, or a *.dc.html file — prints a one-line steer that the harness injects as context. It is silent and non-blocking on every other prompt.

Enable (requires your approval — it changes agent startup config)

Merge into .claude/settings.json (into any existing hooks object):

{
  "hooks": {
    "UserPromptSubmit": [
      {
        "hooks": [
          { "type": "command", "command": "sh \"<skill-root>/hooks/detect-handoff.sh\"" }
        ]
      }
    ]
  }
}

Restart Claude Code afterwards. Verify:

printf '%s' '{"prompt":"… claude.ai/design/p/x?file=Y.dc.html"}' \
  | sh <skill-root>/hooks/detect-handoff.sh   # prints the steer
printf '%s' '{"prompt":"fix a failing test"}' \
  | sh <skill-root>/hooks/detect-handoff.sh   # prints nothing

The skill cannot add this for you: registering a hook modifies agent-behavior config and is blocked by the auto-mode self-modification guard pending explicit user approval. Add the snippet yourself, or ask the operator to add it and approve the change.