Size the work in your own engagement
Take a feature from an engagement you actually run and decompose it until each task is independently reviewable, then state where your current cadence breaks the exit gates.
Decomposition plan
Human effort concentrates at Define and Judge
Every task declares file ownership and the plan states where your real cadence conflicts with the gates.
Supplied engagementMeridian Payments · Idempotent refund submissionheadless sliceStages 6 and 7 leave the fixture and use an engagement you run yourself.
Why this stage exists
The supplied engagement is behind you. The last two stages work on something you own, under the constraints you actually have.
Four to eight tasks per spec is the target. Work that cannot be decomposed that far returns to specification rather than to a longer agent run.
What you start from
Work from these inputs only. Anything not on this list is either a later stage’s concern or a decision you do not own.
- Your engagement
- A real feature you expect to deliver in the next interval.
- The cadence
- Section 08 of the operating process, effort against wall clock across seven stages.
- Your constraints
- The review capacity, interval boundary and approval path you actually have.
Human effort concentrates at Define and Judge
One worked pass, so the shape of the work is visible before you do it on your own.
- Split the feature into tasks of roughly 15 to 30 minutes of agent work and one reviewable diff each.
- Declare file ownership per task so parallel branches cannot collide.
- Draw the dependency graph and mark what can run in parallel.
- Place your human effort against the stages and find where you are under-resourced.
What this turns onBlocked beats guessing. Ambiguity returns to a person, and a cadence that has no room for that is a cadence that will produce unverifiable results.
Do the work
- Decompose one real feature to the task level.
- Mark every task that cannot be reviewed independently and split it again.
- State the one exit gate your current process would fail today.
- Write what you would have to change to stop failing it.
You produce
Decomposition plan
- Task
- Files owned
- Depends on
- Reviewable alone
- Exit gate at risk
Challenge your own result
Your interval is fixed and the feature does not fit. What gives?
Show how this resolves
The scope, not the gates. A slice that cannot be specified, evaluated, countersigned and learned from inside the interval is not ready for outcome delivery, and shipping it anyway converts a planning problem into an evidence problem.
Ready to continue when
- Every task declares file ownership.
- The plan names the exit gate your current process would fail.
What transfersSize work by what can be reviewed, not by what can be built.
Source · 08 Cadence and decomposition