A cross-repository collection of reusable scripts, skill modules, and shared guardrail instructions for AI coding agents, acting as a canonical mirror for multi-project specifications and tooling.
agent-scripts is a cross-repository AI coding agent toolkit and specification collection maintained by Peter Steinberger. Open-sourced under the MIT license, the project is primarily written in Python, TypeScript, JavaScript, and Shell, with no formal release tags published yet.
Core Design: Canonical Mirror Pattern#
The project adopts a "canonical mirror" pattern, implementing a single source of truth for multi-repository behavioral specifications via the AGENTS.MD pointer mechanism. Downstream repositories only need a single pointer line READ ~/Projects/agent-scripts/AGENTS.MD BEFORE ANYTHING to reference the full rule set without copying. Specifications cover commit conventions (Conventional Commits: feat/fix/refactor/build/ci/chore/docs/style/perf/test), code style, PR workflow, CI handling, security policies, and Git operation boundaries. Git safety uses a tiered strategy: by default only git status/diff/log (read-only) are allowed; push operations require explicit user request; destructive operations like reset --hard and clean require explicit user consent. Model preferences record Anthropic Opus 4.5/Sonnet 4.5, OpenAI GPT-5.2, xAI Grok-4.1 Fast, Google Gemini 3 Flash and other latest models.
CLI Tools#
- committer (
scripts/committer): Bash commit helper that precisely stages specified files and enforces non-empty commit messages. - docs-list (
scripts/docs-list.ts): Traverses thedocs/directory, validates front-matter (summary, read_when), and outputs document summary lists. - browser-tools (
scripts/browser-tools.ts): Standalone Chrome DevTools helper inspired by Mario Zechner's "What if you don't need MCP?" article, supporting Chrome launch/check, navigation, JS execution, screenshots, content search, element inspection, and process termination.
Skill Modules#
The skills/ directory contains 20+ reusable agent skill modules, some contributed by @Dimillian:
| Category | Examples |
|---|---|
| Apple Platform | swift-concurrency-expert, swiftui-liquid-glass, swiftui-performance-audit, swiftui-view-refactor, instruments-profiling, native-app-performance |
| AI / Image | openai-image-gen, oracle |
| Web / Frontend | frontend-design, create-cli |
| Search / Conversion | brave-search, markdown-converter, video-transcript-downloader |
| DevOps / Tools | 1password, domain-dns-ops, discord-clawd, openclaw-relay, sonos-debug, xurl, speaking |
Implementation Details#
- Zero-dependency self-contained principle: All scripts avoid tsconfig path aliases and shared source directories, running independently in any project.
- Bun-compiled standalone binaries: TypeScript scripts are compiled into platform-native binaries via
bun build --compile; thebin/directory is not version-controlled. - Sync strategy: Any modifications to shared scripts must first be synced back to this repository, then bidirectionally synced to all downstream repositories, maintaining byte-level consistency.
Installation & Usage#
git clone https://github.com/steipete/agent-scripts.git
Add a pointer reference at the top of the target repository's AGENTS.MD, then compile CLI tools using Bun:
bun build scripts/docs-list.ts --compile --outfile bin/docs-list
bun build scripts/browser-tools.ts --compile --target bun --outfile bin/browser-tools
External Tools Reference#
tools.md documents tools available in the author's environment (bird, sonoscli, peekaboo, clawdis, oracle, mcporter, etc.) as references only, without containing their implementations.
Unconfirmed Information#
- The relationship with OpenAI Codex / Claude Code is convention-based coordination, with no formal integration SDK.
- The agent discovery and loading mechanism for skills modules is not explicitly documented.
- Minimum Bun version requirement for compilation is not specified.