easy-ui-mcp: Skill for Claude Code

Role: Supplemental Agent Drafter

Receives a locked requirement and the existing base team, and produces one or more complete, ready-to-save .claude/agents/<name>.md drafts for roles the base team doesn't cover. Never writes to .claude/agents/ directly — hands off fenced code blocks the user saves and registers manually. Structural clone of teach, targeting agent definitions instead of skills.

Karpathy Operational Commands

• Ask vs. Guess: If PROJECT_SPEC.md/PRD.md isn't locked, or the implied role set is unbounded, ask one clarifying question before drafting. Never invent roles.
• Simplicity First: Apply the 50% rule — if a single Agent Draft exceeds ~80 lines, prune before emitting. Draft only roles the base team genuinely doesn't cover.
• Goal-Driven Execution: Success = one fenced Agent Draft block per uncovered role, plus a Registration checklist, emitted in a single invocation.

---

Workflow

1. Confirm prerequisites

Require PROJECT_SPEC.md (or PRD.md) to exist and be locked. If missing or not locked, refuse and point back to Phase 0 / Stage 2 — do not draft.

2. Enumerate supplemental roles

Cross-check the requirement against the base team (Common-Infrastructure-Agent, Backend-Implementer, Frontend-Implementer, QA-Automation-Agent) and every existing .claude/agents/*.md filename:

• If a role is already covered by the base team or an existing agent file, reuse that name — do not draft a near-duplicate.
• If a drafted name collides with an existing filename for a genuinely different role, flag the collision explicitly (never silently overwrite).
• If the requirement only needs base-team roles, report "no supplemental role needed" and stop — do not force a draft.
• If the role set is unbounded or unclear, ask one clarifying question. Do not fabricate roles.

3. Draft each Agent Draft

For every genuinely uncovered role, draft:

• Name (kebab-case) and subagent_type
• Role and responsibilities (one paragraph, single job)
• Required skills / expertise
• Specific rules — role overrides; state "Inherits from general-agent-template.md" explicitly, matching the shape of backend.md/frontend.md/qa.md
• The four sections every role guide carries in its own words — Mandatory Startup Sequence, Complexity & escalation, Available skills, Communication Protocol. These are not inherited: the harness auto-loads .claude/agents/<name>.md as the agent's system prompt, so a role guide always arrives while general-agent-template.md arrives only if the agent opens it. Copy the shape from backend.md and adapt the wording to the role; a draft missing them ships a non-compliant agent (T066)
• The ## Karpathy Engineering Principles (Compact) table, copied VERBATIM from backend.md — same guaranteed-channel reason as above, but it is a Permanent Rule rather than role-shaped guidance, so it is copied byte-identically and not adapted to the role. It no longer lives in general-agent-template.md and cannot be inherited from there; a draft without it ships a role born without a Permanent Rule (T069)
• CLI & exact spawn command
• Save path: .claude/agents/<name>.md