Claude Code · spec-driven development
A structured pipeline inside Claude Code that transforms business requirements into verified, tested code — using multi-agent review gates at every stage.
Six commands carry a feature from requirements to verified code. The BA owns requirements; the developer drives the rest. Each <feature-id> threads one feature through the whole run.
Gather business requirements into a structured, testable spec with clear acceptance criteria.
Draft the technical plan & task waves, then put it through multi-agent review before you iterate and approve.
Reviewers · Technical · Security · Spec Alignment · GDPR · Pragmatism · Devil's Advocate
Gap analysis — three agents run in parallel, grading findings CRITICAL / IMPORTANT / MINOR.
Agents · Coverage Analyst · Integration Scout · Dependency Analyst
Code generation through a TDD pipeline — tests first, then implementation, in 50–200 line tasks.
Agents · Builder · Test Validator
Verify the implementation provably meets the spec — four agents inspect the finished change.
Checks · Implementation · Constitution · Quality · Tests
Auto-fix failures behind a circuit breaker. Constitutional violations escalate at once; otherwise an Escalation Report hands off to a human.
Runs the full pipeline — require through fix — in a single command.
Express and Fast go through JIRA automation; Standard and Heavy follow the normal process. The split is about how obvious the design is — not how big the change feels.
Repeatable, well-patterned foundational work that follows established conventions.
Small-scope corrections within established patterns — a few hours of work.
Core feature work with novel business logic the design isn't obvious for.
Complex, exploratory, or architectural work needing real engineering judgment.
A well-scoped JIRA ticket labelled express or fast moves to AI Generation Ready; ~15–25 min later a draft PR links back to the ticket. The engineer reviews it like a junior's PR, tests it, refines it — and owns it. Nothing merges without human review.
One sitting following existing patterns → Express or Fast. Non-obvious design decisions → Standard or Heavy. When in doubt, default to Standard — and any agent or engineer can escalate a lane up when reality disagrees.
The shared pipeline, dialled per lane. Automated lanes lean on JIRA + agents with one human sign-off; normal-process lanes add review and approval gates. Human grooming (require) runs in every lane.
| Pipeline stage | L1Express | L2Fast | L3Standard | L4Heavy |
|---|---|---|---|---|
| JIRA automation | auto | auto | read/write | read/write |
| 01.require human | from pattern | from pattern | guided spec | + discovery |
| 02.plan · 6 agents | auto-review | auto-review | 6 reviewers | + design spike |
| 03.analyze · 3 agents | light | light | full | deep |
| 04.build · 2 agents | ● | ● | ● | ● |
| 05.verify · 4 agents | ● | ● | ● | + sec / GDPR |
| 06.fix · max 3 | ● | ● | ● | ● |
| Human gate human | 1 · combined | 1 · combined | 3 gates | 4+ · arch board |
One thread ties the work together — each artifact derives from the one before it, so the acceptance criteria you groom are exactly what gets built and tested. Runs in every lane; only the depth changes.
The intent, captured and gap-checked in require. The source of truth.
Acceptance criteria as Given / When / Then, refined and lane-labelled in refinement.
One case per acceptance criterion, signed off before the build runs.
Executable tests written test-first in build, asserting each case in CI.