# Agent Scripts By **steipete** ยท Skills Open-source collection of reusable agent skills, shared AGENTS.MD rules, and portable helper scripts for Codex and Claude agents, by steipete. - Source: https://github.com/steipete/agent-scripts - Repository: https://github.com/steipete/agent-scripts - Tags: open-source, agent, cli, claude, codex, workflow, self-hosted - Pricing: free - Upvotes: 0 ## Features - SKILL.md-based routing layer: each skill has YAML front matter with name and description for agent matching - AGENTS.MD shared hard rules, symlinked globally into Codex and Claude config directories - Symlink-based skill sharing: repo-owned skills stay canonical in their source repo - scripts/committer: staged commits with non-empty message enforcement and skill validation - scripts/validate-skills: checks all SKILL.md files for valid YAML front matter, runs as a git hook - scripts/browser-tools.ts: Chrome DevTools helper with nav, eval, screenshot, network, and content search commands - scripts/docs-list.ts: walks docs/ and enforces summary and read_when front matter - Progressive disclosure: agents load only the full SKILL.md they choose to use, not all skills at once --- ## Why it matters Every agent-heavy workflow eventually spawns the same problem: rules and skills scatter across repos, drift out of sync, and break silently. Agent Scripts solves this by making one repo the canonical source for shared agent instructions, skills, and helpers, symlinked everywhere rather than copied. ## The big picture The repo organizes four things: `AGENTS.MD` (hard rules loaded before any task), `skills/` (routing-layer workflow definitions), `scripts/` (portable, dependency-light helpers), and `hooks/` (local guardrails like pre-commit skill validation). Downstream repos point to this one with a single pointer line and add only their local overrides below it. ## How it works Each skill is a directory under `skills/` containing a `SKILL.md` file with YAML front matter. The required fields are `name` and `description`. The description is the routing signal: agents match it against the current task and load the full instructions only when they decide to use the skill, a pattern the README calls progressive disclosure. Shared personal skills live as real folders; skills from external repos (such as `discrawl`, `birdclaw`, `slacrawl`) are exposed via tracked relative symlinks so the canonical copy stays in its home repo. ## Zoom in The helper scripts do real work. `scripts/committer` stages exactly the files you list, rejects empty commit messages, and runs skill validation before writing the commit. `scripts/validate-skills` checks every `skills/*/SKILL.md` for valid YAML front matter and required fields, and can run as a git hook via `git config core.hooksPath hooks`. `scripts/browser-tools.ts` is a standalone Chrome DevTools helper that covers navigation, evaluation, screenshots, network inspection, and content search; it compiles to a self-contained binary with Bun. ## Yes, but This repo is built around one person's local workspace conventions (symlinking into `~/Projects/agent-scripts`). Adopting it as-is means accepting those path conventions or adapting them. The README is explicit: keep scripts dependency-free and portable, and never add repo-specific imports or path aliases. ## The bottom line Agent Scripts is the reference implementation for how a single developer (or a small team) can maintain one authoritative set of agent rules and skills and propagate them cleanly across every project via symlinks and pointer-style `AGENTS.MD` files. For anyone running multiple Codex or Claude agent setups, it is the most concrete, working example of that pattern available in the open. ## FAQ ### What is Agent Scripts? Agent Scripts is an open-source GitHub repository by Peter Steinberger that centralizes shared agent instructions, reusable workflow skills, and portable helper scripts for Codex and Claude-style agents. It is designed to be the canonical source for a developer's agent setup, symlinked into every downstream project rather than copied. The repo is MIT licensed and written in Shell, JavaScript, Python, and TypeScript. ### How do I set it up and use it? Clone the repo to ~/Projects/agent-scripts, then create symlinks from ~/.codex/skills and ~/.claude/skills pointing to the skills/ directory, and symlink ~/.claude/CLAUDE.md and ~/.claude/AGENTS.md to AGENTS.MD in the repo. In each downstream project, add a pointer-style AGENTS.MD that reads 'READ ~/Projects/agent-scripts/AGENTS.MD BEFORE ANYTHING (skip if missing)' and place any repo-specific rules below that line. You can enable pre-commit skill validation by running 'git config core.hooksPath hooks' in the repo. ### Is it free and open source? Yes, Agent Scripts is released under the MIT license, which allows use, modification, and redistribution with minimal restrictions. There is no hosted service, subscription, or paid tier; everything runs locally on your machine. ### What is it best for? Agent Scripts is best for developers who run Codex or Claude agents across several repos and want a single, validated source of agent rules and workflow skills rather than duplicating them in each project. The symlink architecture means a fix or new skill in one place propagates everywhere instantly. The included helpers (committer, validate-skills, browser-tools.ts) make it useful as a lightweight local automation layer alongside any agentic workflow. ### How does it compare to other skill or agent-rule approaches? Most agent setups either hard-code rules per repo or rely on a hosted platform's skill marketplace. Agent Scripts takes a middle path: a local, file-based, symlink-propagated system that works with both Codex and Claude without requiring any hosted service. The open agent skills standard (also used by Codex and the Microsoft Agent Framework) defines the same SKILL.md format, so skills authored here are portable to any compatible agent runtime. The tradeoff versus a hosted registry is full local control at the cost of manual distribution. ### What are the limitations? The repo's path conventions assume a ~/Projects/agent-scripts clone location; teams on different layouts must adapt the symlinks manually. Skill distribution is entirely manual, there is no package manager or registry, so sharing skills beyond your own machines means contributors must clone and re-link themselves. The browser-tools.ts script requires Bun to compile to a binary. Very large skill sets may hit context-window limits in agent runtimes that cap the initial skills listing (per the Codex docs, at most 2% of the model context window or 8,000 characters). --- [View on Analog](https://analoghq.ai/steipete/skills/agent-scripts)