Claude Delegate

Claude Delegate

Use Claude Code as a second-opinion or review layer while keeping Codex responsible for scope, edits, validation, and final claims.

Select the execution mode

• Use non-interactive mode for bounded reviews and decision requests:

  claude.cmd -p --no-session-persistence "<focused task>"
  

• Use an interactive terminal when the user wants to type directly, use browser tools, choose a model, or run a longer task:

  claude.cmd
  

  Keep the TTY session alive. Use terminal input only when the user asks Codex to type for them; otherwise leave the prompt available to the user.

Use Fable 5

When Fable 5 is requested, prefer the interactive selector because promotional or subscription-backed model names may not be accepted by the CLI flag:

1. Start claude.cmd.
2. If prompted about Chrome/browser access, ask the user before enabling it.
3. Type /model.
4. Select Fable (the selector identifies it as Fable 5).
5. Choose session-only mode when the request is temporary; choose default mode only when the user asks to make it persistent.
6. Confirm the terminal reports Fable 5 before claiming that Fable 5 reviewed the work.

claude.cmd --model fable-5 and /model fable 5 may fail even when Fable 5 is available in the interactive selector. Do not silently label Opus, Sonnet, or another model as Fable 5. If Fable 5 is unavailable, report that exact limitation and either continue without a Claude review or ask the user to select it interactively.

Write effective delegation prompts

Include:

• the decision or review question;
• the relevant repository or artifact paths;
• constraints and known data boundaries;
• whether Claude may edit files or must remain read-only;
• the exact response format needed;
• validation already performed and what remains unverified.

For code review, request findings first, grouped by severity, with file paths and line references. For product decisions, request one recommendation, alternatives rejected, risks, and acceptance criteria.

Review workflow

1. Inspect the current workspace and diff before delegating.
2. Ask Claude for a narrow, decision-oriented output.
3. Record the model used, review scope, and material conclusions in the PR description or task notes when the work has a review workflow.
4. Apply or reject findings using Codex judgment; do not let Claude's response perform unverified external writes.
5. Run the relevant local tests, builds, and runtime checks independently.
6. Never treat a green Claude response as proof that code, live data, or deployment behavior is correct.

Interactive user handoff

When the user asks to type in Claude:

• launch an interactive TTY session in the requested workspace;
• show the current prompt or confirmation state;
• type only the exact input the user authorizes;
• leave the session active after reaching the Claude prompt;
• report any model, browser, login, or permission prompt that still needs user input.

Safety and truthfulness

• Never pass secrets, cookies, API keys, or repository credentials into a prompt.
• Do not authorize destructive shell commands, external messages, or production changes merely because Claude proposes them.
• Ask before enabling browser access when it could use a signed-in Chrome session.
• Distinguish Claude's recommendation from Codex's implementation and from verified runtime evidence.
• Keep a precise record of which model actually produced the review.