wshobson/agents

The problem

Anyone who uses more than one AI coding tool ends up maintaining the same prompts, reviewer instructions and scaffolding workflows several times, in formats that each tool discovers differently. Switching from Claude Code to Codex or Cursor means re-authoring your subagents and rules from scratch, and there is no shared place to look for a decent set of them. This repository tries to be that place: a single Markdown source for agents, skills and commands, with adapters that emit what each harness expects.

What's different

The usual approach is per-tool: Cursor rules in one directory, Claude subagents in another, a personal dump of prompts in a third. Here the unit of distribution is a plugin that bundles agents, skills and commands together, and installation goes through each harness's own marketplace mechanism where one exists. Two things stand out. First, the breadth: 94 plugins covering Python, JavaScript, review, testing, infrastructure and security, with roughly 200 agents, 185 skills and 105 commands. Second, an explicit capability matrix that documents what each harness can actually load — commands turned into skills, agent formats rewritten, model aliases remapped — rather than pretending every tool is equivalent. The included plugin-eval static checker is a nod toward quality control, though the README itself says its LLM-judge layer is experimental and not validated against human labels.

Who it's for

Developers who already use a coding agent daily and want a curated starting set instead of writing their own subagent prompts, especially teams standardizing across several tools or people moving between Claude Code and Codex. It is also useful as a reference for how to structure plugins if you plan to author your own. Look elsewhere if you only use one tool and are happy with your own rules, if your environment forbids prompt/plugin marketplaces, or if you need audited, benchmarked quality guarantees — the catalog is broad but the README gives no evidence that individual plugins outperform a plain good prompt.

Cost of adoption

MIT licensed, which makes reuse easy. Claude Code and Codex installation is a couple of marketplace commands; OpenCode, Antigravity, Copilot and Pi require cloning the repo plus uv, Python 3.12+ and a make install-* target, which links plugins into your user configuration. Skills-only installs need GitHub CLI 2.90+ or Node/npm. Risks: harness-specific gaps are real (Pi agents need an extension, Codex generated agents go through a separate adapter path), the project is large and fast-changing, and generated artifacts can drift from source. Optional LLM-judge evaluation bills your own Anthropic API key.

Verdict

This is a young but genuinely large catalog — created mid-2025, and it has only recently started showing up on trending monthly boards with roughly 1.8k stars in a month, so interest is rising rather than established. The engineering around multi-harness adaptation is more careful than most prompt collections, and the README is unusually honest about experimental evaluation and per-harness limits. The trade-off is that breadth is not the same as proven usefulness; treat it as a well-organized starting point to fork and prune, not a quality-certified dependency. Install a single plugin, measure whether the agents actually change your output, and keep only what earns its context budget.