waterflow

Settle what the work is before deciding how to build it.

Writing documents agents consume. Use when creating or editing a skill, a reference, or an AGENTS.md / CLAUDE.md.

Implement one slice test-first and close it with recorded evidence.

Review a diff on two separate axes, standards and spec, in parallel subagents. Use when reviewing a branch, a pull request, or work in progress before it lands.

Settle the words and the hard-to-reverse decisions for an area, and record them. Use when a term is being used two ways, a domain concept needs a name, or a decision will look surprising to a future reader.

Investigate a question against primary sources and record what was found. Use when documentation or API facts need gathering, or when reading legwork should be delegated rather than done inline.

Interview the owner in rounds until every open decision is settled. Use before committing to a plan, when facts or decisions are missing, or on any request to stress-test thinking.

Integrate finished work after a final review and a fresh proof.

Build throwaway code that answers one design question. Use when a state model, an interface, or a piece of logic needs checking before anything is committed to it.

Run the named proof for a change and record whether it passed. Use after implementing something, when asked whether a change actually works, or before integrating. Not for writing tests — that is the test atom.

Query what the flow already settled about a subject before deriving it again. Use when starting work on an area, briefing a subagent, resuming after a break, or checking whether a question was already answered.

Choose and confirm where behaviour will be observed, before any test is written. Use when deciding where a test goes, what an interface should expose, or how deep a module should be.

Turn settled work into a proof seam and a set of implementable slices.

Cut work into vertical tracer-bullet slices with blocking edges and publish them. Use when a plan or spec needs breaking into implementable pieces.

Write one failing test at the agreed seam, then make it pass. Use for red-green work, building a feature test-first, or fixing a bug with a reproducing test. To run an existing proof instead, use prove.

Route work by risk. Sets lane, model tier, agent topology, proof, and owner.