Taming prompt sprawl with symlinks and hard rules

What it does This repo collects the shared hard rules, reusable skills, and portable helpers that steer AI agent behavior across Peter’s local projects. It houses global agent instructions in AGENTS.MD, workflow skills under skills/, and small dependency-free scripts for validation, committing, and browser automation. Downstream repositories reference this central copy instead of duplicating agent instructions.

The interesting bit Rather than copying prompts everywhere, the repo uses a pointer-style AGENTS.MD and tracked relative symlinks to expose skills that actually live in other repositories like discrawl or birdclaw. This lets repo-owned skills stay canonical in their own homes while still being discoverable by Codex and Claude through a unified skills/ directory.

Key highlights

• AGENTS.MD provides shared hard rules for Codex/Claude-style agents; global discovery is wired to look here for both toolchains
• skills/ mixes personal skills with symlinked repo-owned skills (birdclaw, discrawl, gog, imsg, slacrawl, wacli, wacrawl)
• scripts/validate-skills enforces YAML front matter (name, description) on every SKILL.md
• scripts/browser-tools.ts offers a standalone Chrome DevTools helper for navigation, screenshots, JS evaluation, and network inspection
• scripts/committer stages files, demands a commit message, and runs skill validation before committing

Caveats

• The entire setup is explicitly tailored to “Peter’s local workspaces,” so paths and symlinks assume a specific sibling-repo layout.
• Several skills are symlinked to paths outside this repo (../../agent-skills/, ../../discrawl/), meaning the full tree only works if those sibling repositories exist on disk.

Verdict Developers juggling multiple repos with AI agents should study the symlink-and-pointer pattern for keeping rules synchronized. Everyone else should treat this as a well-documented personal dotfiles repo rather than a drop-in framework.