ai-skills-agents-context
Health Uyari
- License — License: MIT
- Description — Repository has a description
- Active repo — Last push 0 days ago
- Low visibility — Only 8 GitHub stars
Code Gecti
- Code scan — Scanned 1 files during light audit, no dangerous patterns found
Permissions Gecti
- Permissions — No dangerous permissions requested
Bu listing icin henuz AI raporu yok.
Claude Code agents, skills, and reference guides: security, code-quality and WCAG accessibility auditors, Gemini/GPT/Grok consult agents, and in-depth guides to Agent Skills and MCP.
AI Skills, Agents & Context
⭐ If AI Skills, Agents & Context is useful to you, please star it: click the ☆ Star button at the top right of the GitHub repo page.
Battle-tested Claude Code subagents, skills and long-form context guides ("golden docs"): auditors for security, accessibility and code completeness; consult agents that hand questions to OpenAI, Google Gemini and xAI Grok through their CLIs; and reference guides for writing Agent Skills and MCP servers. Everything is plain Markdown you can drop into ~/.claude/.
Table of Contents
- Overview
- Quick Install
- What's Inside
- Usage
- Requirements
- Security Model
- Repository Layout
- Staying Current
- Contributing
- License
Overview
These files are the working set behind a daily Claude Code setup, published after each change. They solve two problems:
- Reviews that stop halfway. The auditor agents carry long, numbered checklists (secure-coding patterns, WCAG 2.2 success criteria, completeness checks across 15+ languages and IaC tools), so a review covers the same ground every time and reports findings in a fixed, severity-ranked format.
- One model's blind spots. The consult agents let Claude ask another frontier model for a second opinion, live web research or a security audit, while Claude stays the one that writes and commits the code.
Each agent is a single Markdown file with YAML frontmatter (name, description, model, color) followed by its system prompt. The description decides when Claude Code calls the agent, and each one includes <example> blocks of real trigger phrases.
Quick Install
git clone https://github.com/ventz/ai-skills-agents-context
cd ai-skills-agents-context
mkdir -p ~/.claude/agents ~/.claude/skills
cp agents/*.md ~/.claude/agents/ && cp -R skills/* ~/.claude/skills/
Restart Claude Code, then run /agents to confirm they loaded. cp overwrites same-named files, so back up any existing agents first. To install for one project only, copy into that project's .claude/agents/ and .claude/skills/ instead. The consult agents also need their CLIs; see Requirements.
What's Inside
Auditor Agents
Run these after a change, or ask for one by name. Each reports findings only; Claude makes the fixes.
| Agent | What it does | Triggers on |
|---|---|---|
security-auditor |
Severity-ranked security review of code, infrastructure and services: authn/authz, injection, crypto, secrets, supply chain, IaC and cloud IAM, CI/CD, AI/LLM integrations (prompt injection, RAG poisoning), OAuth/OIDC, multi-tenant data access, serverless and edge. | Security-relevant code changes, new Dockerfiles or pipelines, "review this for security" |
accessibility-auditor |
WCAG 2.2 analysis with 208 numbered code patterns, 21 ARIA widget patterns, framework checks (React/Next.js, Vue, Angular, Svelte, Astro), native mobile, email mode, axe-core and ACT rule cross-references, and WCAG-EM 2.0 methodology. | New templates, forms, modals, custom widgets, color or CSS changes |
code-quality-sweeper |
Pre-production completeness audit of an entire codebase: README features vs. real implementation, UI → API → database chains, stubs and dead code, env vars, dependencies, and IaC drift (Terraform, CDK, Helm, Kubernetes, GitHub Actions, CloudFormation, Pulumi and more). Resumable on large repos, with a machine-readable summary for CI gating. | "Make sure nothing is half-implemented before we ship" |
Consult Agents (OpenAI, Google, xAI)
Each wraps a vendor CLI, sends a focused briefing, and returns the other model's analysis to Claude. None of them write final code.
| Agent | Backend | Best for |
|---|---|---|
openai |
codex CLI: GPT-6.2 Sol by default, GPT-6.2 Astra for requested security audits, Daybreak Blue for defensive cybersecurity work |
Architecture and tradeoff analysis, root-cause debugging hypotheses, research that needs reasoning and live web evidence together, security audits |
google |
Antigravity CLI agy (Gemini 3.1 Pro), the successor to the gemini CLI |
Grounded web research, current documentation and best practices, anything about Google products |
xai |
grok CLI: Grok 4.7 on AWS Bedrock, plus an optional live-X lane through a grok.com login |
Cheap, fast, high-volume work: batch classification and extraction, log and corpus triage, tool-calling sweeps over local files |
How the consult agents behave:
- Plan first, then API key. The
openaiagent tries your ChatGPT plan first, falls back to a per-run API key only when the plan doesn't offer the model, and flags every pay-per-token run with a capitalized warning line. - Model routing is explicit. Every invocation pins the model and reasoning effort, so a config change or a new CLI default can't silently switch models.
- Known gaps are written down. For example, Bedrock-hosted Grok has no live web search and will invent results if asked, so the
xaiagent sends live-data questions to thegoogleoropenaiagent instead of answering from training data.
Skills
| Skill | What it does |
|---|---|
documenting-project-context |
Keeps a project's CLAUDE.md short and useful across sessions: what belongs there and what goes in path-scoped rules, skills, hooks or docs/ files, plus compression rules, structure templates, and good and bad examples. Triggers on "update CLAUDE.md", "document this", or when a correction would otherwise be repeated next session. |
Context Guides
Long-form references to load into a session, or to read when building your own.
| Guide | Size | Covers |
|---|---|---|
| Creating Agent Skills | ~2,400 lines | The agentskills.io open standard, Claude Code skill structure, YAML frontmatter fields, invocation control, string substitutions, dynamic context injection, progressive disclosure, security and cross-tool compatibility |
| MCP Development Guide | ~3,300 lines | Python MCP servers and clients on protocol 2026-07-28 and the official SDK v2 (MCPServer): the stateless design, server/discover, Streamable HTTP and stdio transports, subscriptions, multi round-trip requests, caching hints, error codes, and a legacy appendix |
Helper Scripts
| Script | What it does |
|---|---|
bin/daybreak.sh |
A one-shot OpenAI Responses API call under the Daybreak Blue cyber access program, used by the openai agent for defensive security work. It reads the key from the macOS Keychain per run and passes it to curl on stdin, never in argv or the environment. Install: install -m 755 bin/daybreak.sh ~/.claude/agents/bin/ |
Usage
Claude Code picks an agent automatically when a request matches its description; you can also ask for one by name:
> Review the new login endpoint for security issues
> Use the accessibility-auditor on the checkout form
> Ask the openai agent whether we should split this service, and why
> Have the google agent find the current Vertex AI auth docs
> Use the xai agent to bucket these 300 log files by root cause
Tips:
- Auditors work best on a scoped target. Name the files, directory or diff to review; the sweeper is designed for whole repositories and can take a while.
- Consults need a briefing, not a file dump. The agents tell the other model what was already ruled out and what kind of answer to return, and you get back a summary.
- Agents run with
model: claude-opus-5-5. Change themodel:line in a file's frontmatter tosonnet,haikuorinheritto trade depth for cost.
Requirements
The auditor agents and the skill need only Claude Code. Each consult agent needs its CLI on PATH and an account:
| Agent | CLI | Account / auth |
|---|---|---|
openai |
codex (npm install -g @openai/codex) |
A ChatGPT plan login (codex login), and optionally OPENAI_API_KEY for models the plan doesn't serve |
google |
Antigravity CLI agy |
A Google account or a Vertex AI project (set your project ID where the file shows <your-vertex-project>) |
xai |
grok CLI |
An AWS Bedrock API key with Grok access; optionally a grok.com login for the live-X lane |
Exact versions, flags and the gotchas each one was verified against are in the agent files themselves.
Security Model
Anything a consult agent's commands print is sent to that vendor, so the agents are built to keep secrets out of it:
- Empty environment.
codexis launched withenv -i(onlyPATHandHOME), because a single tool that prints its environment would otherwise send every exported credential along. Name-based secret filters don't stop that. - One credential per run. A per-run API key is the only variable added, and never the shell's own key variable.
- Isolated home. The
xaiagent runsgrokwithHOMEpointed at a separate consult directory holding only its config and a symlinked login, so your global instructions don't ride along into the consult. - Read-only by default. Consults run with read-only sandboxes, and the auditors report findings instead of editing code.
If you adapt these agents, keep those rules, and never put credentials, .env contents or auth files in a briefing.
Repository Layout
agents/ Subagent definitions (one .md file each, YAML frontmatter + system prompt)
skills/ Skills (one directory per skill, entry point SKILL.md)
context/ Long-form reference guides
bin/ Helper scripts used by the agents
Staying Current
This repo is updated whenever an agent, skill or guide changes: new model releases, CLI flag changes, and gotchas found in real use. To update:
cd ai-skills-agents-context && git pull
cp agents/*.md ~/.claude/agents/ && cp -R skills/* ~/.claude/skills/
The commit history is the changelog; each commit says what changed and why. Watch the repo (Watch → Custom → Releases or All activity) to hear about updates.
Contributing
Issues and pull requests are welcome, especially reports of an agent missing something it should catch, a CLI flag that has changed, or a model or price that is out of date. Keep each change focused on one agent or guide, and say how you verified it.
License
MIT © Ventz Petkov
Yorumlar (0)
Yorum birakmak icin giris yap.
Yorum birakSonuc bulunamadi