Free · no sign-up
Roast my CLAUDE.md
Paste your CLAUDE.md, AGENTS.md or Cursor rules. Kin scores it out of 100 on what the people who build these agents say matters, quotes the lines that deserve it, and hands you the fixes.
Hand it over. I have read worse. Probably.
Free, no sign-up, 5 a day. Claude reads the file to write the roast; the file is not stored. The roast gets a link that can quote a few lines; you can delete it. What is kept
What Kin checks
Five things every agent file gets scored on
The criteria come from Anthropic’s Best practices for Claude Code and Kyle at HumanLayer’s Writing a good CLAUDE.md. The jokes are Kin’s own.
- 1
Specific: says what the code can’t
Decisions, gotchas and misleading corners, not a tour of the folders or advice any repo would get.
- 2
Runnable: lets it check its work
The exact install, test, typecheck and lint commands, down to running one test.
- 3
Lean: short enough to be read
Every line loads into every session: under about 200 lines, task-specific detail in linked files, style left to the linter.
- 4
Calm: plainly worded
No IMPORTANT, ALL CAPS or YOU MUST. A rule that can’t bend belongs in a hook, not a louder sentence.
- 5
Current: consistent and current
No rules that fight or repeat, no files that are gone, no habits from older models like “think hard”.
The verdict
Nine kinds of agent file
Every roast names the type. Most files are one of the first eight; Kin is still hoping for more of the ninth.
The Novel
Long enough for a sequel. Your agent skims it like terms of service.
The Sticky Note
Three lines and a prayer.
The Hostage Note
NEVER. ALWAYS. IMPORTANT!!! Blink twice if the build is broken.
The Fortune Cookie
Write clean code. Follow best practices. Your agent already knew.
The Wiki Page
Tells the agent everything about the project except what to do.
The Plot Twist
Rule 4 disagrees with rule 11, and the agent has to pick a side.
The Time Capsule
Paths that moved, versions that died, TODOs from another era.
The Junk Drawer
Rules, notes, a changelog and one grocery list.
The Well-Oiled Machine
Short, specific, runnable. Kin came to roast and left taking notes.
What it reads
Any file you give an AI agent
CLAUDE.md
Claude CodeClaude Code loads CLAUDE.md into every session. It can pull in other files with @path imports, a CLAUDE.md in a subfolder loads when the agent works there, and CLAUDE.local.md holds personal notes kept out of git.
AGENTS.md
Codex, Cursor, Amp and othersAGENTS.md is the open format read by Codex, Cursor, Amp, Jules, Gemini CLI, Copilot and others: plain Markdown, no required fields, and the file nearest to the code being edited wins in a monorepo.
Cursor rules
CursorCursor rules live in .cursor/rules/*.mdc with frontmatter (description, globs, alwaysApply) that decides when each rule loads; many small scoped rules beat one big always-on file. .cursorrules is the old single file.
copilot-instructions.md
GitHub Copilot.github/copilot-instructions.md applies to every Copilot chat and agent request in the repository; path-specific .instructions.md files with applyTo globs cover parts of it.
GEMINI.md
Gemini CLIGemini CLI loads GEMINI.md files from the home folder, the project root and subfolders, all joined into the context, and can import other files with @path.
Windsurf rules
WindsurfWindsurf reads .windsurfrules or .windsurf/rules/*.md, with a character limit per rule file and activation modes (always on, by glob, by description, manual).
SKILL.md
Agent skillsA SKILL.md starts with frontmatter: name and description. Only the description is loaded until the skill is used, so it must say what the skill does and when to use it; the body loads on use and can point at scripts and reference files next to it.
instructions file
any AI agentAn instructions file for an AI agent or assistant (a system prompt, .clinerules, CONVENTIONS.md for Aider, .goosehints, a custom GPT or project prompt).
Questions
About the roaster
What does a good CLAUDE.md or AGENTS.md look like?
Short, specific and runnable. It holds what the agent cannot work out from the code: the exact commands to install, test, typecheck and lint (down to one test), the decisions that are not written anywhere else, and the places that mislead. Anthropic suggests keeping CLAUDE.md under about 200 lines; HumanLayer keeps theirs under 60. Detail for one kind of task goes in its own file that the main one points to.
How is the score worked out?
Kin scores five things from 0 to 10: Specific, Runnable, Lean, Calm and Current. Specific and Runnable count a little more, because a file that says nothing about this repository, or gives the agent no way to check its work, fails at the job whatever else it does. The weighted mean becomes a score out of 100 and a grade from S (frame it) to F (charcoal).
Is my file stored?
No. The file goes to Claude once, to write the roast, and is not kept. Anything that looks like a key or token is taken out in your browser before it is sent. What is kept is the roast itself (the scores, the jokes, the few lines it quotes and the fixes) under a random link, so it can be shared. The browser that made it can delete it.
Why does it say ALL CAPS and IMPORTANT are bad?
Current models follow instructions closely, so a rule in capitals gets applied too widely rather than more correctly. Anthropic’s guide says to drop IMPORTANT and YOU MUST and give the reason instead. A rule that must never bend (never push to main, never touch the migrations) is better as a hook or a permission rule, which the agent cannot ignore.
Can I roast a public repository’s file?
Yes. Paste a GitHub link (github.com/owner/repo, or a link to the file) and Kin looks for CLAUDE.md, AGENTS.md, GEMINI.md, .cursorrules, Copilot instructions and similar files in it. Only public files on GitHub are read, and the result names the repository.
Will the fixes work as they are?
They are written from the file alone, so commands and paths Kin cannot know are left as placeholders in angle brackets. The roast ends with one prompt that holds every fix: paste it into the agent that works in your repository and it can check each fix against the real code before it writes it.
Which files does it read?
CLAUDE.md for Claude Code, AGENTS.md for Codex, Cursor, Amp, Jules and others, Cursor rules (.cursorrules and .mdc), .github/copilot-instructions.md, GEMINI.md, Windsurf rules, a SKILL.md, or any system prompt or instructions file for an AI agent.
Who made this?
Briefkin, a board that shows what every AI coding agent you run is doing: working, waiting on you, done. Kin is its mascot. The roaster is free; the criteria come from Anthropic’s Claude Code best practices and HumanLayer’s guide to writing a good CLAUDE.md.
The file is not stored; the roast is, until you delete it. What is kept, in full.
Your agents still need watching.
Briefkin shows every Claude Code, Codex and Cursor session on one board: working, waiting on you, done. Answer them from your phone. Free in the open beta.