Understand the system and the problem before proposing solutions.
No code. No feature plans. Read, verify, ask.
Understand the system, the problem, and the scope. Produce the foundational artifacts that all subsequent work builds on.
Always. No exceptions. Even trivial changes require reading CLAUDE.md and confirming scope.
User requirement, existing code & architecture, CLAUDE.md, project documentation.
Nine artifacts — in order. Numbers 2–4 form the uncertainty triplet.
Lightweight plan: what to read/verify, open questions, draft scope boundaries.
Who has the problem? What does not work today? Why does it matter? Why now?
What we think could be true and could solve the problem; each hypothesis carries a cheapest test and a kill criterion.
Where we might be wrong, top-3 expensive failure modes, counter-evidence we may be ignoring.
What the system is, what it does, who it serves, which components it has.
High-level architecture: layers, modules, communication paths, technology stack.
Architecture Decision Records for all fundamental design decisions.
Catalogue of existing artifacts: code, docs, schemas, conventions, dependencies.
Only after 1–8: confirm scope with the human.
Context is where uncertainty is explicitly catalogued — not assumed away
Concrete answers to: who has the problem, what fails today, why it matters, why now, what counts as "solved".
≥2 testable hypotheses. Each one ships with the cheapest way to confirm or kill it.
≥3 risks rated. Top-3 expensive failure modes called out, plus counter-evidence the team may be ignoring.
Each surfaced via AskUserQuestion — never as prose
Every checkbox must be satisfied before moving to Exploration
Four of the seven cross-cutting roles operate in Context
Process governance, scope confirmation, gate enforcement.
System Spec, problem formulation, initial scope definition.
Architecture document, ADRs, codebase analysis.
Context Inventory, what AI must read, context hierarchy.