Preserve architecture
Tell the agent what to reuse, what not to rewrite, and how far a change should travel.
6 coding agents · one shared source · no sign-up
Which AI agent instruction files does your repo actually need?
Choose Codex, Claude Code, Cursor, Copilot, Gemini CLI, or Devin Desktop (formerly Windsurf). Get the smallest documented file set without copying the same rules across every tool.
1. Choose your tools
The recommendation updates immediately.
Nothing is uploaded. The planner uses only your tool selections.
2. Use the smallest correct file set
4 files for 6 tools
# AI coding agent instruction file plan Selected tools: OpenAI Codex, Claude Code, Cursor, GitHub Copilot, Gemini CLI, Devin Desktop (formerly Windsurf) ## 1. AGENTS.md — canonical shared instructions Keep repository-wide commands, architecture boundaries, safety rules, and the definition of done here. Treat this file as the source of truth. ## 2. CLAUDE.md — Claude Code bridge ```md @AGENTS.md ``` Add Claude-only guidance below the import only when it genuinely differs from the shared rules. ## 3. .github/copilot-instructions.md — Copilot surface bridge ```md # Copilot compatibility Follow the repository-wide rules in the root AGENTS.md. ``` Keep this bridge only for Copilot surfaces that do not load AGENTS.md. Copilot CLI can replace the pointer above with a real relative include: ```md @../AGENTS.md ``` Do not assume that @ file references expand outside Copilot CLI. If a surface neither loads AGENTS.md nor expands the reference, keep the minimum essential repository rules self-contained here and verify the active instructions in that surface. ## 4. GEMINI.md — Gemini CLI bridge ```md @./AGENTS.md ``` Add Gemini-only context below the import only when needed. ## Tool-specific notes - Codex: AGENTS.md is the repository instruction file; no bridge is needed. - Cursor: AGENTS.md currently applies at the project root. Use .cursor/rules/*.mdc when a rule needs directory or file-pattern scope, globs, or another activation mode. - Devin Desktop (formerly Windsurf): AGENTS.md works directly. Use .devin/rules/*.md for cross-cutting rules that need explicit activation behavior; .windsurf/rules/*.md remains a legacy path. - Claude Code: verify the import with /memory after adding CLAUDE.md. - GitHub Copilot: cloud agent, code review, VS Code, and CLI can load AGENTS.md. Keep the .github bridge for surfaces that require repository-wide custom instructions; @ file references are explicitly documented for Copilot CLI only. - Gemini CLI: verify loaded context with /memory show. ## Drift-prevention checklist - Edit shared rules in AGENTS.md first. - Keep bridge files short and tool-specific. - Use scoped native rules only when the scope cannot be expressed by file location. - Re-check loaded instructions after changing filenames, imports, or repository layout.
Next step
Need the canonical instruction file itself?
Create a browser-local first draft, then use the plan above to connect it to the rest of your coding tools.
Live preview
AGENTS.md
# Acme dashboard An internal analytics dashboard for operations teams. ## Start here - Read the files nearest to the requested change before editing. - Follow the existing Next.js, TypeScript, and Tailwind CSS patterns and preserve the current architecture. - Use npm; do not introduce a second package manager or lockfile. - Treat existing user changes as intentional. Never discard or overwrite them. ## Working rules - Keep changes tightly scoped to the request. Avoid speculative refactors. - Run `npm test` after changes and fix failures caused by your work. - For UI work, preserve keyboard access, visible focus, semantic markup, and responsive behavior. - Never expose secrets, credentials, private data, or local environment values. - Explain non-obvious tradeoffs in the final handoff. ## Project-specific constraints Do not change database migrations without explicit approval. ## Definition of done - The requested behavior works end to end. - Relevant checks pass and no new warnings are introduced. - Changed code follows existing naming, formatting, and module boundaries. - The handoff states what changed, what was verified, and any remaining risk.
Generated locally in your browser. Your project details are not uploaded.
New · manifest import
Already have a package.json or pyproject.toml?
Drop in one manifest and get a local, copy-ready AGENTS.md with detected commands—no CLI, upload, or API key.
Why it works
A small file with an outsized effect.
Coding agents are better when the repository defines the boundaries. Repo Agent Kit turns vague preferences into instructions an agent can actually follow.
Define done
Make tests, checks, accessibility, and the final handoff part of the task—not an afterthought.
Protect the repository
Keep user changes, credentials, lockfiles, and migrations away from accidental damage.
One source, six agents
Stop copying the same rules into every tool.
Choose Codex, Claude Code, Cursor, Copilot, Gemini CLI, or Devin Desktop (formerly Windsurf) and get the smallest documented file set with AGENTS.md as the shared source.
New · public repo audit
Audit a GitHub repo—or generate its AGENTS.md.
Check readiness in a shareable report, then create an evidence-backed instruction file without an AI key.
See the 10-repository benchmarkQuestions, answered.
What is an AGENTS.md file?
AGENTS.md is a repository instruction file that tells compatible coding agents how to work in a project: commands, conventions, guardrails, and the definition of done.
Where should I put AGENTS.md?
Place it at the repository root for project-wide guidance. Larger repositories can add more specific files inside subdirectories when local rules need to override the root guidance.
Does this work with Claude Code and Cursor?
Yes. Choose your agent and Repo Agent Kit changes the filename and, where needed, the file structure or frontmatter.
Can I check a public GitHub repository?
Yes. The repository audit checks public files without GitHub sign-in and creates a shareable readiness report. Private repositories are never requested.
Are my project details sent to a server?
No. The current generator builds and downloads the file entirely in your browser.