codebase-discovery

原文涵盖的主题

• User Input
• Purpose
• Core principle
• Roles
• Modes
• Graceful degradation (optional inputs)
• Working state (resumable, no hooks)
• Freshness check (staleness detection, no hooks)

You MUST consider the user input before proceeding (if not empty). It identifies the codebase (or subsystem) to analyse and, optionally, the mode. -- Reverse-engineer enough business and domain knowledge out of an existing codebase to onboard a new team member — human or AI — and to give AI harness tooling (e.g. Spec Kit) the context it needs before any specification or change work begins. The output is a small, lean set of onboarding documents under docs/, not a comprehensive knowledge base. Each document is written so it can be linked from a CLAUDE.md / AGENTS.md without consuming an unreasonable amount of context. This skill is the orchestrator. It runs five phases, each defined in its own playbook under playbooks/. Read and follow the relevant playbook at each phase. -- The code is ground truth for what the system does. Only people hold the why. So the method is: mine the code first to form evidence-backed hypotheses, then spend the human's time validating intent and explaining, not re-deriving mechanics. Existing docs (README, CLAUDE.md, AGENTS.md, wikis) are a valuable starting point, but because documentation naturally drifts from code over time, the source code is the sourc

这里只展示节选，请以完整原文和当前要求为准。