collaborating-with-claude_skill

Overview

This skill lets you collaborate with Claude Code via a lightweight bridge script to consult, review, and propose code changes while keeping Codex as the primary implementer. It supports multi-turn sessions using SESSION_ID and is optimized for low-token, file/line-based handoffs. The workflow enforces safe, verifiable output such as Unified Diff patches that you apply locally.

How this skill works

You run a bridge script that communicates with Claude Code and returns structured JSON, including success, SESSION_ID, agent messages, and optional errors. The bridge supports streaming JSON, multi-turn sessions, model selection, and working-directory control so Claude can reference files by path and line ranges rather than requiring pasted snippets. For code edits, ask Claude to produce Unified Diff Patch ONLY and apply the patch locally after verification.

When to use it

• You want a second opinion on design tradeoffs, edge cases, or missing tests.
• You need Claude to propose or review a Unified Diff rather than directly modify files.
• You want multi-turn back-and-forth while implementing locally and keeping context via SESSION_ID.
• You prefer file/line references to reduce token usage and protect sensitive data.
• You need a reviewer to spot risks, readability issues, or missing tests before merging.

Best practices

• Always run the bridge script (do not invoke Claude directly) so session and output handling stays consistent.
• Provide entry files and approximate line numbers instead of pasting large code blocks.
• Request OUTPUT: Unified Diff Patch ONLY when asking for code changes and forbid direct modifications.
• Capture and reuse SESSION_ID for follow-up prompts to maintain conversation-aware context.
• Set a sensible default timeout (10 minutes when using the Codex runner) and prefer the sonnet model for routine tasks; use opus only for complex requests.
• Keep a short Collaboration State Capsule updated to summarize goal, SESSION_ID, files/lines handed off, last ask, and next ask.

Example use cases

• Start a session to analyze a failing unit test and then request a minimal Unified Diff fix while preserving conversation context.
• Ask Claude to review a proposed patch and list edge cases or missing tests before you apply it.
• Use file and line references to have Claude suggest refactors for a large module without pasting files into the prompt.
• Run iterative design discussions: get tradeoff analysis, then request a sample patch, then ask for tests to validate the change.