SebastienElet/agent-skills

Claude Developer

Overview

Prepare a launch command and one self-contained prompt for a manual Claude Code handoff. The user runs the command, pastes the prompt, and returns with the result when ready. Implementation and correction prompts compose claude-development, which owns implementation, checks and any requested Codex review.

Usage

$claude-developer <task or returned result> — prepare an implementation or correction handoff. Direct implementation requests without Claude stay with the current agent.

Steps

1. Inspect repository instructions and the minimum read-only context needed for the handoff. Record unresolved facts instead of guessing.
2. Identify the intended existing worktree from the task context; verify its absolute path and branch with git worktree list --porcelain. Use the current worktree when the task targets it. If the intended worktree is missing or ambiguous, ask for its path before issuing a launch command; do not invent a path or create a checkout.
3. Write one prompt with the outcome, verified worktree path and branch (or detached HEAD), relevant facts, constraints, acceptance criteria, allowed scope, expected verification, and delivery report. Tell Claude to inspect before editing, preserve unrelated changes, and surface conflicts. Include the user's authorization limits; reserve consequential actions for the user unless explicitly authorized.
4. For an implementation or correction handoff, explicitly require Claude to activate claude-development on the authorized task. It owns implementation, checks and any explicitly requested independent Codex review before delivery. Forward the original scope and bounded local correction authority, preserve stricter user limits, and require the report's candidate, dispositions, evidence and incomplete outcomes. Do not duplicate that finishing procedure in the prompt. Planning-only and review-only handoffs do not authorize development. Do not restart it inside a code-review-fix correction or infer open-PR repair or publication authority from this handoff.
5. Return two fenced blocks in order: a sh block with the copyable one-line command cd <shell-quoted absolute worktree path> && claude, then a text block with the prompt to paste into that session. Substitute the verified path, quoting shell metacharacters safely. Repeat both blocks for every replacement or corrective prompt, then stop.
6. On a returned result, review only supplied evidence and explicitly authorized local state. Produce a corrective handoff only when the user requests another iteration; an unexecuted prompt revision replaces the previous handoff.

Gotchas

• Starting in the wrong checkout — a bare claude uses the terminal's current directory; supply the verified absolute path and && so a failed cd prevents launch.
• Losing the manual handoff — executing the command or starting another iteration removes the user's control; supply the two blocks and wait.
• Trusting a returned summary — it may omit defects or failed checks; inspect the available diff and evidence before making a verification claim.
• Inventing a review requirement — a handoff can impose unrequested Codex work; compose claude-development once and preserve only the reviews and correction authority actually requested.

Constraints

• Activate only for requests explicitly involving Claude Code, a manual handoff, or its returned result; intercept automatic Claude delegation as a manual handoff.
• During preparation, never invoke Claude, an automation wrapper, or another implementation agent; never create a branch or checkout, edit files, or validate the implementation.
• Never authorize commits, pushes, merges, deletion, publication, or permission bypasses unless the user's current request explicitly permits them.
• Produce at most one prompt per response and wait for the user before every subsequent iteration.
• Never require an unrequested final Codex review, treat a prompted review as observed evidence, or bypass the user’s repair limits.
• This is advisory policy, not a technical execution barrier. Never claim Claude's result is verified without reviewing the named evidence in its stated environment.