sruthik27/creating-claude-md

This repository is an Agent Skill that helps a coding agent generate a short CLAUDE.md or AGENTS.md file for a project from scratch. A CLAUDE.md file is the project memory that tools like Claude Code or Codex read at the start of a session, so the agent knows what the repo is for and which rules to follow. The skill packages the steps needed to produce a good first version. The README says the skill works in two passes. First, the agent does an autonomous scan of the repo, looking at the package manifests, the folder layout, linter configs, git history, existing docs, and any configured MCP servers. Second, it runs a focused six-question interview with the user to collect the things only the user can answer: the WHY of the project, hidden gotchas, what is out of scope, when approval is needed, the testing philosophy, and links to external docs like Confluence pages. The target output is a CLAUDE.md under 80 lines with a hard cap of 120 lines. The README explains the reasoning. If a line would not prevent a specific agent mistake, the skill deletes it. The author argues that bloat in CLAUDE.md causes coding harnesses to give less weight to every instruction in the file, so the line cap is treated as a forcing function for high signal content. The README states this skill is for repos that have no CLAUDE.md, or where the existing file is too far gone to salvage. It is not meant for auditing or incrementally improving a working CLAUDE.md. Installation for Claude Code is a git clone into ~/.claude/skills. For Codex or other agent runtimes, you point the agent at the SKILL.md file. The project is released under MIT.