// the find
wshobson/agents
Multi-harness agentic plugin marketplace for Claude Code, Codex CLI, Cursor, OpenCode, GitHub Copilot, and Gemini CLI
A large collection of Markdown-based prompt templates (agents, skills, commands) packaged as installable plugins for Claude Code and several other AI coding tools. The core idea is a single source-of-truth directory that gets transformed into harness-native artifacts for each supported tool. Aimed at developers who want pre-built AI workflow patterns without writing their own system prompts.
The multi-harness adapter pattern is the genuinely interesting part — one Markdown source emits idiomatic output for six different tools rather than duplicating content per-tool. The tiered model strategy (Fable 5 for long-horizon work down to Haiku for fast ops) is a sensible cost-control approach that most teams figure out the hard way. The plugin-eval framework with static + LLM-judge + Monte Carlo layers is a rare attempt to actually measure prompt quality rather than just ship it. Plugin isolation (installing one plugin loads only its context) avoids the context-window bloat problem that kills naive 'load everything' setups.
The scale numbers (192 agents, 156 skills) are marketing, not engineering — more Markdown files does not mean more working behavior, and there's no evidence the individual agents have been validated to produce consistently better output than a good system prompt. The build pipeline depends on `make generate` + gitignored output trees, which means Gemini and OpenCode users must clone and run local tooling just to install, creating a maintenance cliff when upstream harness specs change. The 36k stars against 3974 forks and a repo this young is a suspicious ratio, suggesting star inflation. There's also no versioning story for plugins — if an agent changes behavior between your install and the next `make generate`, you find out by noticing worse output.