AI Agent 配置管理工具

English | 中文 | Changelog

Awesome Agent Config

One repository for Claude Code and Codex: global instructions, coding rules, plugins, shared skills, status lines and correction memory. Your existing agent detects the platform, explains the options and installs your choices. The same conversation works on macOS, Linux, Windows and WSL, with commands checked against the local client.

Showcase

• adversarial-review skill — worked example

Quick Start

The active release line is Mizoreww/awesome-agent-config on main. Upgrading from the previous repository name or development branch? Follow the v4 migration notes.

Open this checkout in Claude or Codex, or share the URL of the repository page you are reading along with the request below. Keep its branch/ref when sharing a branch page.

Read INSTALL.md from the checkout or repository page I shared, using the same branch/revision as its README. Configure the agent I am talking to. Detect my OS, client and existing configuration. Directly in this conversation, list every option supported by this agent by category with continuous numbers, a useful description, author recommendations and installed status. Prefer native multi-select questions when available; otherwise let me choose multiple numbers or names in chat. Keep the choices in the conversation; do not generate a separate Markdown document or report. Install and verify my choices, preserving my customizations.

The agent shows only your target agent's supported options directly in the conversation, with useful descriptions and recommendations. If the client offers a real multi-select question tool, choose options by category there; otherwise reply with multiple numbers, names or a description of what you want. The agent respects the tool's actual limits and reuses choices you already made. A plugin bundle gets one number and lists its members; capabilities included in that bundle are not installed twice. Author recommendations follow the recorded legacy main/Codex installer defaults and subsequent changes, with explicit mappings: 20 Claude and 18 Codex recommendations in this release. Recommendations do not select items for you.

For later changes, say “add asd-ste100”, “update my previous selections” or “remove storage-analyzer”. Existing choices are reused; omitting an installed item never uninstalls it. Both agents use edit-config for configuration queries, additions, edits, removals, repairs and updates. It tracks this repository’s main branch. Queries are read-only; if the skill is absent, the global instructions link to the same workflow.

Native plugin/MCP commands are preferred; the agent chooses source installation when the documented scope or client requires it. You do not need to choose between npx, plugins and skill copies yourself. Codex setup disables external-agent auto-import so installations follow your selections. Windows and WSL are detected and configured separately; App/CLI installations sharing a home are reused.

For Codex, the agent checks the available OpenAI official/curated directory, then upstream Codex packages and verified compatible plugins. The plugin guide records exact routes and limitations; official availability depends on the account and client. Existing equivalent capabilities are reused.

Installation workflow · Claude operations · Codex operations · Full catalogue and bundle members

Catalogue

The tables retain the original categories and merge the Claude and Codex capabilities. Third-party originals are installed from their upstream sources; this repository stores author-owned and intentionally customized skills with attribution. Handoff belongs to the Matt bundle only. — means this repository does not offer that item for the agent. The agent hides those entries when presenting your choices. A platform cell marked ★ identifies an author recommendation for that agent. Exact routes, stable IDs and recommendation markers live in catalog.md.

Core

Language Rules

Review

Workflow

Integrations

Design & Content

Slides

Memory & Lifestyle

Storage

Academic Research

MCP Servers

Complete bundle membership is listed in catalog.md. Selected source revisions and adaptations are in sources.md. Storage analyzer modifications are documented in UPSTREAM.md and submitted as khazix-skills#50. Context7 and Playwright appear under Integrations and are not duplicated as separate MCP choices.

Directory Structure

.
├── README.md / README.zh-CN.md   # User guide and category tables
├── AGENTS.md / CLAUDE.md         # Instructions for working in this repository
├── INSTALL.md                   # Agent-guided installation and updates
├── MAINTAIN.md                  # Agent workflow for changing this repository
├── catalog.md                   # IDs, support, recommendations and routes
├── skills/                      # Author-owned and customized skill sources
├── platforms/
│   ├── claude/                  # Claude instructions, lessons, rules, hooks, skills
│   ├── codex/                   # Codex instructions, lessons, settings, skills
│   └── sources.md               # External revisions, members and adaptations
├── scripts/                     # Protected file operations and focused helpers
├── lessons.md                   # This repository's correction history
├── docs/                        # Showcases, spec and historical migration notes
└── install.sh / install.ps1     # Compatibility notices pointing to INSTALL.md

Key Mechanisms

• Agent-guided choices — complete numbered lists, explanations and a record of the user's selections. Plugins and MCP servers are managed through the target agent's native tools; custom files use protected copies and partial merges.
• Independent memory — Claude uses its own global lessons.md plus project memory/MEMORY.md; Codex uses its own global lessons.md plus project-root lessons.md. Templates and real histories remain separate. Only missing global logs are seeded.
• Rules and status lines — Claude has one writing rule and independent Python / TypeScript / Go rules; the gradient status line shows model, directory, venv, Git, context and usage. Codex uses its native footer and subagent capabilities; this repository no longer installs custom role presets.
• Configuration management — edit-config follows main and records the actual revision. Conflicting source policies require an explicit migration choice; user selections and customizations are preserved. Installation and maintenance need no legacy Claude/Codex branches.
• Scoped changes — preserve user edits, credentials, hooks and memory databases. Backups and file ownership support safe updates and explicit removals. ResearchStudio Idea/Reel and PPT Master prepare complete source only; runtime dependencies are handled on first use.

Settings Defaults

These are values from the current templates, applied only when the corresponding item is selected. Existing overrides are preserved and differences reported. Base settings, permissions, status lines and lessons are separate choices; Codex's auto-import setting is an installation prerequisite. The agent checks local version support before applying a setting.

Actual settings and per-item operations: Claude · Codex. Current platform validation limits are recorded in migration notes.

Customization

You can ask the agent to maintain the repository itself. It follows MAINTAIN.md, updates the complete payload and installation route, and keeps the catalogue and both READMEs consistent.

For installed configuration queries and changes, use edit-config, which follows INSTALL.md for modifications. The root AGENTS.md, CLAUDE.md and lessons.md belong to this repository; deployable global files live under each platform's templates.

Acknowledgements

• Claude Code in Action — Anthropic Academy's official course
• Working for 10 Claude Codes by Hu Yuanming — multi-instance patterns
• Harness Engineering by OpenAI
• Anthropic Engineering / OpenAI Engineering
• Claude Code Best Practice by shanraisshan
• Claude How To by luongnv89

License

This repository is MIT. Bundled and fetched third-party components retain their own licenses. lieflat-charts uses PolyForm Noncommercial 1.0.0, for noncommercial use only; it is fetched only when selected and is not redistributed here.