CLAUDE.md

• When a repository wants a dedicated CLAUDE.md file for Claude Code
• When you want starter CLAUDE.md context to use through Oi API or MCP
• When Claude-specific instructions should be separated from shared AGENTS.md guidance
• When the repository wants AGENTS.md as the canonical file and CLAUDE.md as a lightweight Anthropic-facing shim

# CLAUDE.md ## Purpose This file adds Claude Code specific guidance for this repository. Treat AGENTS.md as the canonical shared instruction file and use CLAUDE.md only for Anthropic-facing behavior that is genuinely Claude-specific. ## Read Order 1. User request and linked ticket 2. Product or architecture docs 3. AGENTS.md 4. CLAUDE.md 5. README or incidental docs Do not duplicate AGENTS.md unless Claude Code needs materially different handling. ## When To Use CLAUDE.md Use this file for Claude Code specific execution guidance such as: - how Claude should treat project memory and repository instructions - Claude-specific planning, response-shape, or tool-use preferences - workflow details that matter because Claude Code reads CLAUDE.md directly - constraints that should not be imposed on other agents in AGENTS.md If a rule applies equally well to Codex, ChatGPT, Gemini, Cursor, or other runtimes, it belongs in AGENTS.md instead. ## Claude-Specific Working Style - Keep repository-specific rules in AGENTS.md and Claude-specific additions here - Prefer concise, execution-oriented instructions over generic AI policy language - Use structure when it improves clarity, not by default - Avoid duplicating large repo context blocks that are already covered elsewhere - Keep local overrides and personal memory separate from shared repository guidance ## Task Execution - Inspect relevant files before proposing or making changes - Follow the real code path instead of guessed architecture - Preserve repository conventions unless the task explicitly changes them - Keep edits minimal, scoped, and easy to review - State assumptions when context is incomplete - For larger tasks, keep the plan tied to observable repository facts and acceptance criteria ## Output Preferences - Lead with the answer, change, or finding rather than long framing - Keep explanations concise by default and expand only when the task needs it - Distinguish clearly between confirmed facts, inference, and open questions - Make tradeoffs explicit when recommending a path ## Code And Review Guidance - Prefer concrete code-level reasoning over abstract best-practice language - When reviewing, prioritize correctness, regression risk, missing validation, and unclear behavior - When proposing edits, describe them in terms of actual files, flows, or contracts - Do not suggest broad cleanup unless it materially improves the requested outcome ## Validation - Run the smallest useful validation for the changed area - Report what was validated and what remains unverified - Surface risks, assumptions, and follow-up work clearly - Match validation depth to the risk of the change ## Keep In Sync When AGENTS.md and CLAUDE.md overlap: - Put shared repository rules in AGENTS.md - Put Claude Code specific differences here - Remove duplication when it stops being useful - Update both when the workflow meaningfully changes

• User request and linked ticket
• Product or architecture docs
• AGENTS.md
• CLAUDE.md
• README or incidental docs

Markdown