| name | onboard |
| description | Produce a structured "get up to speed on this whole project" orientation — vision, architecture, where things live, conventions, how to do common tasks, current state — grounded in the real code/config with pointers. Optionally writes/updates docs/ONBOARDING.md. Use when asked to onboard, give a project overview, or "how does this project work / where do I start". Read-only except the optional doc write. |
| argument-hint | ["--write to also produce docs/ONBOARDING.md"] |
| allowed-tools | Read, Glob, Grep, Bash, Edit, Write, Skill, mcp__serena__initial_instructions, mcp__serena__get_symbols_overview, mcp__serena__find_symbol, mcp__serena__find_referencing_symbols, mcp__serena__find_declaration, mcp__serena__find_implementations, mcp__serena__search_for_pattern, mcp__serena__get_diagnostics_for_file, mcp__serena__list_dir, mcp__serena__find_file |
Serena MCP is mandatory for code (read-only here). First call mcp__serena__initial_instructions to load the Serena tool manual, then use Serena for ALL code reading / searching / navigation — prefer symbol navigation (get_symbols_overview / find_symbol / find_referencing_symbols) over whole-file reads.
Onboard to the project
Produce a coherent orientation to the WHOLE project, grounded in the real files. Obey the source-of-truth
authority hierarchy in CLAUDE.md (code → SQL → telemetry → docs → AI output) — verify against the code,
treat docs as leads. Documentarian, not critic (describe what exists; don't critique/fix unless asked).
Method
- Reconnaissance via Glob/Grep — NOT Read-everything. Detect manifests (
*.csproj, package.json, …),
framework/config, entry points, the top 2 levels of the tree (exclude node_modules/bin/obj/.next/.git),
tooling (lint/format/CI/Docker), test layout, and the .claude/ governance (rules, skills, agents, hooks, workflows).
- Vision first. Lead with the product context (
docs/product-overview.md + the injected product-context):
the why / who / moat — per "lead with the product vision, not the tech".
- Architecture map. Stack, layout (monorepo
apps/web + apps/api), layers + dependency direction, API
style, key directories' purpose, and one traced request lifecycle (reuse explain-codebase on a
representative path).
- Conventions from the code itself — naming, error handling (
Result<T> / ProblemDetails), async discipline,
tenancy. Verify against the rules but confirm in code: if config says X but the code does Y, trust the code.
- Rank, don't dump. Surface the most-referenced / central files and the authoritative rules/skills — not an
exhaustive listing.
- Common tasks — verified, not assumed. The real build/test/run commands (from docs/config), and the operating
model: write a plan (
implementation-plan-format.md) → /run-impl-loop; add an endpoint (add-endpoint skill),
a domain entity (add-domain-entity), a command/query (cqrs-kommand), tests (write-tests).
- "Where to look" map. Table: goal → location (incl. the rules/skills/agents that govern that area).
- Flag unknowns / drift explicitly — never guess; if doc↔code drift is obvious, note it (and suggest
/docs-standards-sync).
- Optional artifact (
--write): write/refresh docs/ONBOARDING.md. If it (or CLAUDE.md) already exists,
read it first, enhance, preserve existing instructions, and call out exactly what changed — never
blind-overwrite. Keep it tight.
Output
A skimmable orientation in that order (vision → architecture → conventions → key files → common tasks →
where-to-look → open questions), citing real files. Optionally the docs/ONBOARDING.md artifact.
Pre-scaffold note: until product code exists, onboarding centers on the vision + the governance layer
(the rules/skills/agents/workflows/hooks + the operating model). Onboard to that, grounded in the actual files.