lightarktech How you cut the work decides how it fails. Layer-cuts ("first the database, then the backend, then the UI") mean nothing is demoable until everything is done — and they breed marathon agents: we watched one chew through 8 layer-tasks across hundreds of turns, getting slower and dumber as its context bloated.
For the full ticket template and the worked wide-change pattern, see ticket-template.md.
The rule
A ticket is a vertical slice: one narrow feature path cut through every layer — data, logic, interface, test — that a user (or the founder) can see working on its own.
- Sized to one sitting: the whole ticket fits in a single fresh AI context window. If it doesn't fit, it's two tickets.
- Independently verifiable: each finished ticket is demoable — the founder can try it and say "yes, that's it" (see
decision-wiring, seat 3). - Carries its own acceptance checklist: each ticket lists the checkable items that mean "done", and names the seam its tests will hit (the public interface, agreed up front). A ticket without a checklist is a wish.
- Declares its blockers: each ticket names which tickets must land first. No blockers = can start now, in parallel.
- One ticket, one agent, one fresh context. The ticket boundary is the agent contract boundary (see
dispatch-economics).
The exception
A wide mechanical change (rename something used everywhere) can't slice vertically — one edit breaks a thousand call sites. Sequence it as expand → migrate in batches → contract: add the new form alongside the old, move callers over batch by batch with everything staying green, delete the old form last.
Success criteria
Every ticket ends in something demoable. No agent ever inherits a second ticket in the same context. The founder saw progress this week, not "80% done" for three weeks.