My current AGENTS.md and collection of skills for coding agents.
Collaboration-first, not autonomy-first. The goal is to amplify my thinking, not replace it. I do the thinking about what to do, the agent accelerates the how.
Clone the repo, then symlink (or copy) into the coding agent's config directory:
git clone https://github.com/spieseba/agent-config.git
# Claude Code
ln -s /path/to/AGENTS.md ~/.claude/CLAUDE.md
ln -s /path/to/skills ~/.claude/skills
# Codex
ln -s /path/to/AGENTS.md ~/.codex/AGENTS.md
ln -s /path/to/skills ~/.codex/skills
# Antigravity CLI
ln -s /path/to/AGENTS.md ~/.gemini/GEMINI.md
ln -s /path/to/skills ~/.gemini/skills
# Mistral Vibe CLI
ln -s /path/to/AGENTS.md ~/.vibe/AGENTS.md
ln -s /path/to/skills ~/.vibe/skillsMy AGENTS.md is inspired by forrestchang/andrej-karpathy-skills @ 2c60614, which itself is derived from Andrej Karpathy's observations on coding-agent pitfalls. It materializes the following four principles: Think Before Coding, Simplicity First, Surgical Changes, Goal-Driven Execution.
grill-me— Get extensively interviewed about a plan or design until every branch of the decision tree is resolved.- This skill is inspired by mattpocock/skills and obra/superpowers.
- Key idea: force decisions to be surfaced before code is written.
- Note: Grill isn't one-shot. When you've been grilled and you're now coding, you'll hit a fork that wasn't covered. Re-invoke.
claude/statusline-command.sh is my Claude Code status line: user@host, cwd, and model on line 1; context usage plus 5h/7d rate-limit bars and session count on line 2. POSIX sh; needs jq on PATH.
Claude Code wires this up via settings.json, which I don't sync (it holds machine-specific permissions and env vars). So symlink the script, then point at it manually:
ln -s /path/to/claude/statusline-command.sh ~/.claude/statusline-command.sh// merge into ~/.claude/settings.json
{ "statusLine": { "type": "command", "command": "~/.claude/statusline-command.sh" } }