Decision 0066 · Accepted
Specs sized by delivery,
not length
Spec 061 stopped specs growing past 3,000 words with a guideline and a deterministic failure. 0022 required the first slice to test the riskiest assumption.
- Status
- Accepted
- Built from:
- da88c48
Status: Accepted (2026-11-03). Spec: 065. Amends 0022 (slice 1 need not test the riskiest assumption or ship instrumentation), 0023 (contracts are required where agents or other systems depend on a capability, not in every spec), 0060 (item 1: slicing is for work with several outcomes, by outcome first), 0062 (item 3, which restates it), and the length rule of spec 061.
Context
Spec 061 stopped specs growing past 3,000 words with a guideline and a deterministic failure. 0022 required the first slice to test the riskiest assumption. Both served lean discipline, but a word count is a proxy, and a developer who understands their MVP has no use for a forced experiment.
Decision
- Four levels. Product vision and core journeys (
PRODUCT.md, may be broad); the first useful release (the outcome that makes it worth using); specs (independently deliverable increments); tasks (executable slices). Only specs and tasks are spec-file content. to-specdecides the split. Default one. Split only for independent delivery, excessive scope, a consequential dependency, or later investment that should wait on earlier learning. Never for length or task count. A later spec is acandidatestub.- Concise, not capped. No word limit, no length failure. Specs stay short by linking and by leaving implementation to
build. - What, not how. Specs name outcomes, constraints, and consequential decisions; components and file layout appear only as established constraints. Contracts stay where agents or other systems depend on a capability.
- Tasks follow outcomes. Dependencies and integration cost set the count; there is no target.
buildslices only work with several outcomes, and parallel work remains the host's call under 0062. Its test is elapsed time saved against briefing, integrating, and reviewing: this deliberately restores the "when it pays" judgment 0060 rejected, because the host, not a Flow rule, now makes the call. Areas no longer need to be coarse.
Alternatives
- Raise the limit. Rejected: any number invites the same wrong split.
- Warn instead of fail. Rejected: a warning nobody acts on is noise; review already flags padding.
- One spec per MVP. Rejected: it hides what can ship and learn early.
- A new PRD or roadmap document. Rejected:
PRODUCT.mdalready holds the vision.
Consequences
- Specs can drift longer; the guard is now judgment and review, not a script. If they do, revisit with evidence.
explore's stress test stays for new products, and is shorter for a settled vision.