Create CLAUDE.md context files for AI coding agents

The skill creates a folder-specific CLAUDE.md for a target subdirectory, plus an AGENTS.md symlink pointing at it, so any future coding agent working in that folder (Claude Code, Codex, or others) has the folder-scoped context the project's global CLAUDE.md doesn't cover. It follows a seven-step process: confirm the target folder and check it actually deserves a file, read every markdown, config, and key source file in the folder in full without skimming, draft a bullet-list outline grouped by candidate sections - Product/Purpose, Avatar/Audience, Essential Files, Constraints, Conventions, Locked Decisions, Context, How to work with the user, Marketing Angles, Top Insights - and let the user react before writing anything. After the user iterates on the draft (the skill re-reads their edits and flags contradictions, typos, or wrong categorization), it writes the file, using ## headers for the approved sections, bullets over prose, and @relative/path/file.md import syntax for cross-folder references. It closes by creating the symlink and verifying it:

cd <folder> && ln -s CLAUDE.md AGENTS.md

It only commits (git add -A, a Day N:-style message, then push) if the user explicitly asks.

Use it when a subdirectory carries context a future agent session needs and the global CLAUDE.md doesn't already give it: active work in progress, folder-specific conventions, or locked decisions that shouldn't be re-litigated. If it's unsure whether a folder actually qualifies, the skill asks the user rather than guessing. Don't use it on a folder of static reference files - agents can read those on demand, so a dedicated CLAUDE.md adds nothing there. The skill also refuses to duplicate what the global CLAUDE.md already states (personality, dates, ports, and similar project-wide facts stay out of the folder file), and it explicitly bars file trees, directory dumps, or stack details the code already shows - only decisions, rules, and context belong in the output. It also keeps hard MUST NOT rules in a Constraints section separate from "usually do X" patterns in Conventions, on the stated reasoning that splitting the two improves adherence.

Input is a target folder path (absolute, under ~/Documents/code/workspace/ in the reference setup) plus the full contents of every file inside it, read directly rather than skimmed - for large tldraw or Vite subprojects the read step is scoped to package.json, src/App.tsx, one representative module file, and the folder's own module-details.md-style files instead of every file indiscriminately. Output is two files: <folder>/CLAUDE.md, opening with a one-line purpose header (and, for a subdirectory whose parent already has its own file, the line "Apply root CLAUDE.md first, then this file."), organized under ## section headers the user approved, with heavy reference docs annotated **Read when:** so agents don't load them every session; and <folder>/AGENTS.md, created as a symlink to CLAUDE.md rather than a copy, so the two files never drift apart.

It targets agent harnesses that read a CLAUDE.md/AGENTS.md pair for local context, named explicitly as Claude Code and Codex. A companion background reference for the same workflow lives at library/claude-code/claude-and-agents-md.md in the same collection.

Developers and teams running AI coding agents (Claude Code, Codex, and similar) across a multi-folder codebase, who want each subfolder to hand agents its own conventions, constraints, and decisions rather than relying on one global CLAUDE.md alone. The skill is adapted from davidondrej/skills, so adopters should verify their own local paths, tools, credentials, and agent features before acting on its output, and anything beyond writing the two files - remote access, scheduling, browser automation, or other file-changing workflows - needs explicit user approval and a confirmed target environment first.