onboard

Build a mental model of a codebase you're new to — broad first, then narrowing — the way the Claude Code onboarding recipe prescribes. Read only; never modify anything. Delegate the heavy reading to **Explore** subagents so their file dumps stay out of this conversation and only the summaries return. ## Orientation (injected — already in front of you) Top level: ```! ls -1 "${CLAUDE_PROJECT_DIR:-.}" 2>/dev/null | head -40 ``` Manifests / entry points present at the root: ```! ls -1 "${CLAUDE_PROJECT_DIR:-.}" 2>/dev/null | grep -iE '^(package\.json|pyproject\.toml|go\.mod|cargo\.toml|gemfile|composer\.json|pom\.xml|build\.gradle|makefile|dockerfile|readme)' || echo "(no standard manifest at root — this may be a subdir or an unusual layout)" ``` ## Step 1: Dispatch the tour Launch **Explore** subagents (in parallel where independent) to answer these, starting broad and narrowing. If `$ARGUMENTS` names a focus area, weight every prompt toward it. 1. **Overview** — what this project is, who uses it, the top-level layout and what each major directory holds. Point the agent at the README and manifests first. 2. **Architecture** — the main patterns, the entry points, how one request (or CLI invocation) flows end to end, and the module boundaries (what must not depend on what). 3. **Data models** — the core domain types/entities/tables and their relationships, and where they're defined. 4. **Auth & security** — how authentication and authorization work, where secrets/config load from, and the trust boundaries. 5. **Conventions** — naming, file layout, test style and framework, error-handling and logging patterns, and the real build/test/lint commands from the manifests. Each Explore prompt must ask for a tight summary with `file:path` pointers, NOT pasted file contents. ## Step 2: Synthesize the brief Merge the returns into one scannable onboarding brief — nothing pasted verbatim, everything a pointer: ``` # Onboarding: <project> **What it is**: <2-3 sentences> **Stack**: <languages, frameworks, key libraries> **Architecture**: <the 3-5 load-bearing facts + entry points, as file:path> **Data model**: <core entities and where they live> **Auth & boundaries**: <how auth works, where secrets load> **Conventions**: <naming, tests, error handling — the house style> **Run it**: <build / test / lint commands> **Glossary**: <project-specific terms an outsider wouldn't know> **Start here**: <the 2-3 files to read first for the focus area> ``` If the repo is empty or trivial (no source beyond config), say so plainly instead of padding the brief. ## Rules - Strictly read-only. This skill maps the code; it never changes it. - Summaries and `file:line` pointers only — never paste large files or diffs into the brief. - Prefer several focused Explore subagents over reading everything in the main context; keeping exploration out of this conversation is the whole point. - Don't invent structure you didn't verify. If an area (e.g. auth) doesn't exist, say "none found" — don't fabricate one.