Curated, CI-enforced skills and subagents for AI coding agents. One-line install. Zero dependencies. Every file validated — structure, links, security properties, and helper scripts — on every commit, on Linux and macOS.
/plugin marketplace add ayaqen/Agents-Scripts
/plugin install agents-scripts@agents-scripts
Works natively in Claude Code (every skill doubles as a /slash-command), via npx skills add ayaqen/Agents-Scripts for the 70+ runtimes in the skills-CLI ecosystem, and via git clone && ./scripts/sync-skills for Codex and manual setups.
The agent-skills ecosystem has an inversion problem: distribution scaled, quality didn't. Zero-curation registries list hundreds of thousands of skills with no vetting; a community audit found roughly three quarters of sampled community skills scoring below 60/100 — most failing silently; and security researchers have demonstrated real prompt-injection attacks delivered through third-party skills and hooks. Meanwhile every installed skill's metadata occupies context on every turn, so bloated catalogs literally cost tokens. (Sources and methodology: docs/evaluation.md.)
This repo takes the opposite bet — a small catalog where every claim is enforced by CI, not asserted by a README:
| Guarantee | How it's enforced |
|---|---|
| One-line install | Claude Code plugin packaging, checked by validate-plugin in CI |
| Spec-level validation | Front-matter allow-lists, enforced body structure, dead-link checks, shellcheck, 60+ self-tests — on Ubuntu and macOS |
| Static security properties | No dynamic shell preprocessing (!`) in any skill (validator-banned); front-matter keys allow-listed, so skills can't smuggle hooks or shell config; zero dependencies; helpers make no network calls |
| Token discipline | Routing descriptions hard-capped at 200 chars, bodies at 120 lines; a deliberately small catalog instead of a firehose |
| Portability | AGENTS.md + SKILL.md open formats; bash 3.2 + python3 stdlib only; a fresh clone passes CI anywhere |
17 workflow skills — transferable engineering practice, not personal tool wrappers. Each is model-invocable and a slash command:
| Skill | Use it when |
|---|---|
| repo-onboarding | Mapping an unfamiliar codebase before changing it |
| debug-loop | Hypothesis-driven debugging of failing code or tests |
| regression-fix | Fixing a bug with a failing test written first |
| safe-refactor | Behavior-preserving refactors, verified at each step |
| self-review | Reviewing your own diff before committing |
| pr-review | Reviewing someone else's pull request before merge |
| security-review | Security pass over a diff, feature, or dependency |
| ci-green | Driving a failing CI pipeline back to green |
| release-checklist | Cutting and verifying a release |
| dependency-vet | Evaluating a dependency before adding it |
| session-handoff | Writing a handoff so the next session can continue |
| context-budget | Working in large codebases without drowning context |
| model-tiering | Frontier-model orchestration with cheaper executor subagents |
| prompt-triage | Debugging LLM prompt / agent misbehavior |
| eval-design | Designing evals for an LLM-powered feature |
| tool-design | Designing tools and MCP servers agents use correctly |
| skill-author | Creating or updating skills in this repo |
5 tiered subagents — the model-tiering pattern, shipped as installable agents. Every agent body ends with a validator-enforced ## Output contract, because an orchestrator only ever sees the final report:
| Agent | Tier | Role |
|---|---|---|
| code-reviewer | sonnet | Read-only diff review; blocking/non-blocking findings with file:line anchors |
| security-auditor | inherit | Traces untrusted input to sinks; confirms exploitability before reporting |
| bug-hunter | inherit | Reproduces and root-causes; returns proven cause plus dead-hypothesis list |
| test-writer | sonnet | Failing-first regression tests matching your suite's conventions |
| mechanical-editor | haiku | Executes exact-spec edits; stops and reports on any spec mismatch |
git clone https://github.com/ayaqen/Agents-Scripts.git
cd Agents-Scripts
./scripts/doctor # environment + repo health check
./scripts/sync-skills # link skills into ~/.claude/skills and ~/.codex/skills
git config core.hooksPath hooks # optional: block commits that break validationCreate a new skill (scaffolded, validated, safe with any description text):
./scripts/new-skill my-workflow "Short trigger phrase for routing."AGENTS.md Shared hard rules for every agent session
CLAUDE.md Pointer file for harnesses that only read CLAUDE.md
skills/<name>/ One skill per directory: SKILL.md + optional scripts/
agents/ Subagent definitions with enforced output contracts
.claude-plugin/ Plugin + marketplace manifests (one-line install)
scripts/ validate-{skills,docs,agents,plugin,links},
new-skill, sync-skills, committer, doctor
templates/skill/ Scaffold used by new-skill
hooks/ pre-commit guardrail
docs/ Architecture, authoring guide, install, evaluation,
design decisions
tests/ Fixture-based self-tests for all tooling (run in CI)
- Portable or it doesn't ship. No personal paths, no symlinks into sibling repos, no macOS-only assumptions.
- One toolchain. bash (3.2-compatible) + python3 stdlib. Nothing to install before the guardrails run.
- Validation is enforced, not suggested. The same validators run in
committer, the pre-commit hook, and CI — a file that doesn't parse cannot merge. - Curated beats big. Fewer, broader skills route better than many narrow ones, and cost less context. Extend an existing skill before adding a neighbor.
- Front matter is deliberately dumb. Simple
key: "value"lines, allow-listed keys. Universally parseable, and structurally incapable of smuggling hook/shell configuration. - Local beats global. Machine-specific rules live in untracked
AGENTS.local.md, never in shared files.
- Evaluation: the landscape, demand signals, and roadmap
- Architecture and enforcement chain
- Installation (all paths)
- Skill authoring contract
- Design decisions vs. the original agent-scripts
See CONTRIBUTING.md. The bar: portable content, validated structure, real earned knowledge in every Pitfalls section.