AI Agents Library

Production-Ready AI Agent Prompts & Skills

---

What's Inside

Agent skills are portable instruction sets that teach AI agents how to execute domain work well — each is a directory with a SKILL.md entry point (YAML frontmatter + routing), progressive-disclosure references, and self-contained scripts and data.

Two things, usable independently:

• frameworks/shared-skills/ — 140 agent skills following the Agent Skills specification. Drop into ~/.claude/skills/.
• custom-gpt/ — 28 Custom GPT system prompts for ChatGPT, written to fit the GPT builder's 8000-character Instructions limit.

Quick Start

git clone https://github.com/vasilyu1983/AI-Agents-public.git
cd AI-Agents-public

# One skill
ln -s "$PWD/frameworks/shared-skills/skills/ai-rag" ~/.claude/skills/ai-rag

# All of them
for s in frameworks/shared-skills/skills/*/; do
  ln -s "$PWD/$s" ~/.claude/skills/"$(basename "$s")"
done

Symlinks mean edits take effect immediately — no re-sync step.

For a Custom GPT: open custom-gpt/<category>/<agent>/01_*.md, paste into the GPT builder's Instructions field, and attach the sibling 02_sources-*.json as Knowledge if present.

Agent Skills

Family	Count	Covers
software-*	34	Backend, frontend, iOS, Android, desktop, security, payments, performance
ai-*	21	RAG, evals, prompt engineering, pretraining, post-training, inference, MLOps
qa-*	17	Test strategy, debugging, resilience, observability, accessibility, Playwright
foundations-*	16	Queueing, control, information, game, decision, reliability theory for systems
ai-coding-agents-*	13	Coding-agent runtime internals: sandboxing, permissions, sessions, tools
dev-*	10	Git workflow, API design, dependency management, code graphs, planning
agents-*	6	Hooks, MCP servers, skill authoring, swarm orchestration, repo memory
data-*	5	Lakehouse, streaming, SQL optimization, analytics engineering
ops-*	4	DevOps platform, incident response, cost optimization, CI/CD
document-*	4	docx, pdf, pptx, xlsx
docs-*	3	Docs-as-code, AI-ready PRDs, note retrieval
research-*	3	arXiv triage, GitHub mining, research method extraction
product-*	2	Product management, help centers
gamedev-*	2	Godot, Roblox

Two areas are unusually deep and harder to find elsewhere: coding-agent runtime internals (how agent CLIs actually implement sandboxing, permissions, and session resume) and systems-theory foundations (queueing and control theory applied to real design decisions, not as abstract math).

Full breakdown: frameworks/shared-skills/README.md.

How a skill is structured

<skill-name>/
├── SKILL.md      # Entry point — frontmatter + routing. Kept small.
├── references/   # Depth, loaded only when the task needs it
├── scripts/      # Deterministic helpers — no model call needed
├── assets/       # Templates and output contracts
└── data/         # sources.json and curated data

SKILL.md stays cheap enough to hold in context; references/ carry the detail and cost nothing until read.

Custom GPT Agents

28 agents across six categories:

Category	Count
Productivity	8
Lifestyle	7
Programming	5
Education	3
Research & Analysis	3
Writing	2

Each folder holds 01_*.md (the system prompt) and, where useful, a curated sources JSON.

Repository Layout

AI-Agents-public/
├── custom-gpt/                      # 28 Custom GPT prompts by category
├── frameworks/
│   └── shared-skills/skills/        # 140 agent skills
├── CONTRIBUTING.md
└── LICENSE

How This Library Is Built

These skills are the public subset of a larger private library that is run as an engineered system, not a prompt collection. Every skill here passed through that system's quality machinery:

• A routing engine — 8 domain routers dispatch 500+ decision scenarios, each an ordered skill chain with an explicit gate and a named failure mode. Skills are written to compose, not just to exist.
• A generated knowledge graph with drift gates — the catalog (nodes, ownership, ordered scenario edges) is exported from the markdown and CI-checked against it, so prose and structure provably cannot diverge. A pre-tool hook blocks confabulated skill names at invocation time.
• Adversarial review — scenario content is periodically challenged by parallel expert-panel agent runs (consistency, currency, decision value), and fixes are verified by retrieval smoke tests before merging.
• Static security screening — third-party skills are scanned for prompt injection, hidden Unicode, and exfiltration sinks before they enter the library (taxonomy adapted from NVIDIA/SkillSpector).
• Freshness discipline — volatile facts (framework versions, model pricing) live in per-skill data/ files refreshed by script, and prose is gated against its own data to stop version rot.

Standards & Compatibility

• Specification: skills follow the Agent Skills specification — portable directories, YAML frontmatter, progressive disclosure.
• Platforms: Claude Code (~/.claude/skills) and Codex CLI (~/.agents/skills); the same directories work in both.
• Self-containment: every skill works detached from this repo — own data/, own scripts/, symlink-safe path resolution — so a single copied folder is a working install.

Scope

This is a curated public subset of a larger private library. Project-specific, client-scoped, marketing, startup, and legal skills are not published, and skills depending on internal libraries are excluded because they cannot be used outside their origin repo.

Contributing

Issues and PRs welcome — see CONTRIBUTING.md. Useful contributions: correcting something that has gone stale, adding a reference that saved you time, or reporting a skill that misfires.

License

MIT — see LICENSE. Use freely, including commercially.