ai-bootstrap

Personal bootstrap for a new machine: shared agent instructions, the Claude Code statusline, personal skills shared by Claude Code and Codex, and the plugin setup to replay.

Everything here is meant to be linked from the home directory rather than copied, so that a git pull in this repository immediately updates the live configuration.

1. AGENTS.md

AGENTS.md holds the general instructions that every agent must follow (writing style, commit rules, engineering standards).

The goal is to keep a single source of truth and expose it under the two filenames different tools look for: ~/.claude/CLAUDE.md for Claude Code, and ~/AGENTS.md for agents that follow the AGENTS.md convention.

mkdir -p ~/.claude
ln -sfn "$PWD/AGENTS.md" ~/.claude/CLAUDE.md
ln -sfn "$PWD/AGENTS.md" ~/AGENTS.md

Run the commands from the root of this repository, since $PWD is used to build an absolute symlink target.

If ~/.claude/CLAUDE.md or ~/AGENTS.md already exist as regular files, back them up first: ln -sfn replaces them without asking.

The content of AGENTS.md was inspired by Kun Chen (talk on YouTube).

2. statusline-command.sh

statusline-command.sh renders the Claude Code statusline: the first line mirrors the colored PS1 from ~/.bashrc, the second line shows model, context window usage, session cost and duration, and the Claude.ai 5h/7d rate limits.

ln -sfn "$PWD/statusline-command.sh" ~/.claude/statusline-command.sh
chmod +x "$PWD/statusline-command.sh"

The script depends on jq, which must be available in $PATH.

The symlink alone does not enable the statusline: Claude Code must be pointed at it, either with /statusline or by adding this entry to ~/.claude/settings.json.

{
  "statusLine": {
    "type": "command",
    "command": "~/.claude/statusline-command.sh"
  }
}

3. Plugin bootstrap

claude-plugins-bootstrap.md documents the marketplaces, the plugins, and the language server binaries that make up the working setup.

Open it and run the commands you actually want on this machine: the /plugin install ... slash commands go into a Claude Code session, the npm install -g ... commands go into a shell.

Language server plugins (*-lsp) do not ship their binary, so each one has to be paired with its npm install -g counterpart, otherwise /plugin reports Executable not found in $PATH.

See claude-plugins-bootstrap.md for the full list and the exact order.

4. Skills

skills/ holds the personal skills shared by Claude Code and Codex, one directory per skill, each with its own SKILL.md.

Both tools use the same skill format and discover user-level skills in ~/.claude/skills/<name>/SKILL.md and ~/.codex/skills/<name>/SKILL.md respectively. Each skill directory is linked individually instead of linking the whole skills/ folder.

That way both user-level skill directories can still host skills that are not meant to be versioned here, without them ending up in this repository.

mkdir -p ~/.claude/skills ~/.codex/skills
for skill in "$PWD"/skills/*/; do
  skill_name="$(basename "$skill")"
  ln -sfn "${skill%/}" ~/.claude/skills/"$skill_name"
  ln -sfn "${skill%/}" ~/.codex/skills/"$skill_name"
done

Run the loop again after adding a new skill to the repository, since a fresh directory needs its own symlink. Restart Claude Code and Codex after installing or changing a skill so that both tools reload their skill catalog.

5. Tools, skills, and plugins tried

Ponytail

Ponytail is a skill and plugin that guides coding agents toward the smallest correct solution, prioritizing YAGNI, existing code, the standard library, and native platform features. Tested with Codex in full mode.

License

Released under the MIT License.