// the find
msitarzewski/agency-agents
A complete AI agency at your fingertips - From frontend wizards to Reddit community ninjas, from whimsy injectors to reality checkers. Each agent is a specialized expert with personality, processes, and proven deliverables.
A collection of 200+ markdown files defining AI agent personas — each one a CLAUDE.md or system prompt you drop into Claude Code, Cursor, Copilot, or a dozen other tools to get a 'specialized' assistant. The accompanying install script copies the right files to the right config directories. Aimed at developers who want opinionated starting points rather than generic 'you are a helpful assistant' system prompts.
The install script actually handles multi-tool targeting well — one flag to pick Claude Code vs Cursor vs Gemini CLI vs eight others, with dry-run support and a subset-by-division option that's useful when a tool has agent count limits. The native desktop app (separate repo) is a genuinely good distribution mechanism: no clone, auto-updates, GUI browsing of the roster. The agent files themselves tend to be more specific than typical 'act as a senior developer' prompts — the network engineer file lists actual platforms (Cisco IOS-XE, Juniper Junos, Palo Alto PAN-OS), which makes the output noticeably less generic. CI lints agents for structural conformance, so the quality floor is consistent.
136k stars on a repo of markdown files is a social media event, not a technical achievement — treat the star count as marketing signal rather than quality signal. The agent definitions have no versioning strategy: if Claude's behavior changes in a future model, there's no way to know which agents silently degraded. The repo has grown by accumulation — 200+ agents with heavy China-market coverage (Douyin, Kuaishou, Baidu, Xiaohongshu, WeChat, Weibo, Bilibili, Taobao) that most Western dev teams will never touch, and no way to filter by relevance without reading the full README table. There's also a real risk of prompt drift: agents define communication style and deliverables in prose, but nothing validates that the actual LLM output matches the spec — the 'Reality Checker' agent checking the others is circular.