About this project

Jules Prompts is a curated library of machine-readable task prompts, Agent Skills, and planted-failure fixtures designed for coding agents. The name is historical; the instructions are harness-agnostic and do not depend on any specific product like Jules, Claude Code, Codex, or Cursor. The repository is organized into several layers: - `_prompts/`: Canonical procedure text, including a website, JSON index, and MCP slash commands. - `skills/`: The same text formatted as Agent Skills, generated from `_prompts/`, ready to copy into `.claude/skills/` or `.agents/skills/`. - `fixtures/`: Miniature repositories with planted defects to demonstrate how skills fail. - `plugin/`: The entire library as an installable bundle with manifest, skills, and slash commands. - `library.json`: A machine-readable list of every procedure, its category, and tier. All generated outputs are checked byte-for-byte in CI to ensure they stay in sync with the source. Adding a new output format is a single entry in the `TARGETS` list of `scripts/emit.py`. Procedures are classified as `core` unless marked `status: legacy`. Core procedures address the failures agents actually have, while legacy procedures are general-purpose task text from 2025 that remains useful but is not what makes this library unique. The library includes a standing doctrine for `AGENTS.md` in `harness/AGENTS.md` to ensure it fires when no skill is selected. Getting started guides are provided in `PROMPTS_GUIDE.md` (explains each prompt and recommended sequence) and `ENVIRONMENT_SETUP.md` (preparing a repository for agent cloning, installation, and testing). Usage options include: - As Agent Skills: Copy skill directories into `.claude/skills/` and paste `harness/AGENTS.md` into the project's `AGENTS.md`. - As an MCP server: Configure the server in `mcpServers` to expose prompts as slash commands in Claude Code, Claude Desktop, VS Code/Copilot Chat, Windsurf, and Zed. The server reads the repository live, and placeholders like `<PR_URL_OR_DIFF_RANGE>` are exposed as arguments. - For humans: Copy the body of a prompt file after the YAML front matter and paste it into the agent's instruction input. - For agents: Fetch `https://jules-prompts.wecanuseai.com/prompts.json`, select a prompt by title/description/category, and fetch the rendered prompt from its URL or read `_prompts/<slug>.md`. - Against fixtures: Use `python scripts/score_fixture.py fixtures/unfailable-tests path/to/REPORT.md` to score a fixture, with verdicts of holds/broken/skipped and a coverage line. Contributing guidelines emphasize adding prompts only for recurring tasks not already covered, keeping YAML front matter aligned, writing harness-agnostic instructions, regenerating skills, updating guides and workflow, adding fixtures for failure-catching prompts, and passing integrity checks.