メニュー
Guide to creating CLI tools and scripts that augment Claude Code. Use when building bin/ executables, automation scripts, hook handlers, or tooling that abstracts repeated agent workflows.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
Structure Claude prompts for clarity and better results using roles, explicit instructions, context, positive framing, and strategic organization. Use when crafting prompts for complex tasks, long documents, tool workflows, or code generation.
Pause agent execution to have the human validate decisions, choose between options, answer freetext, or comment on a document — via the `hl` CLI. Every interaction is a kickoff that returns a job handle immediately; collect the human's answer later with `hl job result`. Use for material design decisions, approval gates, picks between meaningful alternatives, and markdown doc review. Not for trivial yes/no confirmations the agent should decide itself.
Validate technical designs against requirements for coverage, feasibility, and architectural soundness. Use after creating a design.
Validate implementation plans against requirements and design. Use after creating a plan to ensure full coverage and no ambiguities.
Validate requirements documents for EARS compliance, completeness, and consistency. Use after creating requirements.
Design CLIs for humans and LLM agents — subcommand shape, output streams, exit codes, JSON modes, TTY-aware color, structured errors, mutation safety. Use when building or refactoring a CLI, adding machine-readable output, writing --help, deciding stdout vs stderr, or making a tool agent-friendly.
| name | scripts-and-tooling |
| description | Guide to creating CLI tools and scripts that augment Claude Code. Use when building bin/ executables, automation scripts, hook handlers, or tooling that abstracts repeated agent workflows. |
| user-invocable | false |
Scripts abstract away multi-step sequences, complex computation, and ceremony that the agent would otherwise do manually each time. Only the script's output enters context — complex logic costs zero tokens, and the script persists across sessions and context resets.
| Location | Mechanism | Use Case |
|---|---|---|
bin/ on PATH | Agent calls via Bash | General-purpose CLI tools |
scripts/ | Agent calls via Bash | Project-specific automation |
hooks/ scripts | Called by hooks.json | Lifecycle handlers (guards, formatters, loggers) |
skills/*/scripts/ | Bundled with SKILL.md | Deterministic computation a skill needs |
.mcp.json + MCP servers | Claude calls as native tools | API wrappers, database access |
Self-documenting for agents — --help must be comprehensive enough that an agent never needs to look elsewhere. It's the canonical source of truth, not external docs.
Every output is a prompt — the agent's next action depends on your script's output. Design both success and failure output to guide the agent forward, not just report status.
Familiar interfaces — model your CLI after tools the agent already knows (pytest, docker, kubectl). The agent can infer behavior from convention without reading docs.
No interactive input — agents can't respond to prompts. Use flags/args/env vars instead.
Script output is the primary interface between your tool and the agent. A silent success or cryptic error is a dead end.
Success output — confirm what happened, echo IDs/paths the agent needs next, suggest next steps:
# bad
OK
# good
Created deployment deploy-a1b2c3d4 (env: staging, 3 services)
Next:
Check status: deploy status deploy-a1b2c3d4
View logs: deploy logs deploy-a1b2c3d4
Roll back: deploy rollback deploy-a1b2c3d4
Error output — three parts: what went wrong, how to fix it, what to do next:
# bad
Error: connection refused
# good
ERROR: Cannot connect to database at localhost:5432 (connection refused)
Fix: Ensure postgres is running:
brew services start postgresql
# or: docker start postgres-dev
Then retry: ./migrate --target production
Structured output — emit JSON when the agent will consume the result programmatically. Raw human-readable text is fine for notification scripts or when the agent just needs to relay information to the user.