command-creator

Create Claude Code commands — user-invocable workflows that orchestrate agents, interview users, and delegate complex work through structured phases.

---

Core Process

1. Classify Command Scope

Determine structure before drafting:

• How many phases? Simple commands (1-2 phases) vs orchestration commands (5+ phases).
• Does the command need subagents, or can it run entirely in the main context with Read/Glob/Grep/Bash? Default to main context; only add Agent delegation when the work must run in a fresh context or with a different model.
• Interactive (AskUserQuestion gates) or autonomous (runs end-to-end without prompts)?
• What approval gates are needed before destructive or irreversible actions?

2. Research Existing Commands

Use Glob + Read directly to find similar commands in the target project:

• Glob **/commands/*.md, read 2-3 representative files
• Note phase structure, frontmatter fields, and whether they use subagents or direct tools

3. Dedup Audit

Read ${CLAUDE_SKILL_DIR}/../prompt-writer/references/cc-dedup-guide.md before drafting.

Key dedup rules for commands:

• Commands receive CLAUDE.md automatically — do not re-state its global directives.
• Do not repeat Intent Gate, Delegation Check, or barrier semantics — these load every message.
• Do not re-explain agent model routing rules already in global CLAUDE.md.
• Commands specify WHAT to do (which agents, which phases) not HOW CC should behave.

4. Draft Command

Follow phase-based structure from references/command-patterns.md. Each phase uses:

## Phase N: Name

**Goal**: One sentence — what this phase achieves.

**Actions**:

1. First action
2. Second action
3. Third action (with sub-steps if needed)

Rules:

• Phases must be sequential — later phases depend on earlier ones.
• Insert AskUserQuestion approval gates before destructive actions.
• Delegate heavy work to agents — commands orchestrate, agents execute.
• Reference $ARGUMENTS for user-supplied input.
• Use ${CLAUDE_PLUGIN_ROOT} for paths to bundled templates.

5. Frontmatter

---
description: "Short, action-oriented. What does this command do? ≤250 chars."
argument-hint: "[expected-input]"
effort: low | medium | high
---

Notes:

• Do not declare allowed-tools, disallowedTools, or model. CC ignores them at runtime for user-level plugin components; tool access flows from session permissions.
• Always include AskUserQuestion and Agent if the command interviews users or delegates to agents.
• effort signals expected token budget: low (single task, <5 phases), medium (multi-agent, 5-7 phases), high (full orchestration, 7+ phases).
• Do not add model to command frontmatter — commands inherit the session model. Only agents have model routing.

6. Review

Present draft. Verify before finalizing:

• Phases sequential? No phase assumes work from a later phase.
• Approval gates present before destructive or irreversible actions?
• Agent delegation clear? Each Agent call has TASK + EXPECTED OUTCOME + MUST DO + MUST NOT DO + CONTEXT.
• Error handling covers: missing input, agent failure, no changes found, user cancellation.
• $ARGUMENTS referenced at the right phase (usually Phase 1)?
• No ineffective frontmatter declared (model, allowed-tools, disallowedTools)?

---