claude-code

name	claude-code
description	Delegate complex coding tasks, refactoring, and codebase analysis to Anthropic's Claude Code CLI.
when_to_use	When you need to delegate code writing, bug fixing, test generation, codebase refactoring, or code analysis to Claude Code CLI.
allowed-tools	Bash

Claude Code Integration Skill

This skill guides you in delegating complex software engineering and codebase manipulation tasks to the external Claude Code CLI (claude) tool.

YOU ARE BOUND BY THIS CONTRACT

When using this skill to run Claude Code, you are delegating execution to a powerful sub-agent. Follow these instructions precisely to ensure clean execution, proper session management, and context sharing.

---

Technical Specifications & CLI Options

• -p, --print: Runs Claude Code in non-interactive (headless) mode, prints the result, and exits. This flag is mandatory for automated delegation.
• --permission-mode bypassPermissions: Bypasses all interactive tool permission prompts, enabling full autonomy.
• --session-id <uuid>: Starts or continues a conversation using a specific UUID.

---

Step-by-Step Execution Protocol

Step 1: Directory & Context Verification

1. Ensure you run the Bash tool in the project root directory (which is your current working directory).
2. Claude Code executes in a separate process and does not share your in-memory conversation history. You must pass all target file paths (relative to project root), logs, and database schemas explicitly in the prompt.

Step 2: Session ID Management (In-Memory / Task-Scoped Only)

Session IDs must only be reused within the context of the same active task. Do not share sessions across different tasks.

1. Initial Run: Start the first run of the task without a session ID:

   claude -p "$ARGUMENTS" --permission-mode bypassPermissions
   

2. Capture Session ID: Inspect the output of the command. Look for the printed session ID/UUID. Keep this ID in your memory/conversation history for the duration of this task.
3. Follow-up Runs (Same Task Only): If you need to make consecutive runs or follow-up corrections for the same task, pass the captured session ID:

   claude -p "Now fix the failing test you just observed." --session-id <UUID> --permission-mode bypassPermissions
   

4. Task Completion: Once the overall task is finished, discard the session ID. Do not store it to any files.

Step 3: Parse and Summarize Output

1. Extract the final answers, test outcomes, or file modifications from the CLI's stdout.
2. Summarize the changes and report them back to the calling context.

---

Troubleshooting & Guardrails

• Stuck CLI: If execution hangs, it is likely waiting for user input. Ensure --permission-mode bypassPermissions is set.
• Git State: If Claude Code makes modifications, run git status or git diff after execution to review the changes.