How to run it

Free sample

Context management, time budget, going backwards, and what each gate is for.

pipeline/00-overview.md · 2.4 KB

pipeline/00-overview.md

# How to run the pipeline

## The shape

Nine stages, two gates. The gates are the only places a human has to think
hard; everything else is mechanical once the decisions are made.

## Context management

Stages 01–05 can share one session. They are all about the same question
(what are we building) and the accumulated understanding helps.

**Start a fresh session at stage 06.** Paste in the spec and the plan.
Nothing else. The exploratory conversation from stages 01–03 is now debris:
it contains rejected options, earlier phrasings of the requirement, and
half-formed ideas the model cannot distinguish from decisions.

**Start another fresh session at stage 07.** Self-review inside the
implementing context is close to worthless — the model has just spent an
hour convincing itself the implementation is right. Give the review a clean
context with the spec and the diff, and nothing about how the diff came to be.

## Time budget

For a task that would take a competent developer a day:

    01 Intake        2 min
    02 Spec          10 min, mostly you reading
    03 Risk pass     5 min
    GATE 1           however long the decisions take — this is the real work
    04 Plan          5 min
    05 Scaffold      10 min
    GATE 2           5 min reading
    06 Implement     the bulk
    07 Self-review   10 min
    08 Backfill      10 min
    09 Writeup       3 min

If stages 01–05 take longer than the implementation, either the task is
genuinely ambiguous (good — you found that cheaply) or it is too small for
the pipeline.

## Going backwards

The pipeline is not a ratchet.

- Stage 03 finds a risk that changes the requirement → back to 02
- Stage 05 cannot express the shape cleanly → the spec is wrong, back to 02
- Stage 06 hits something the plan did not anticipate → back to 04, not
  forward with an improvisation
- Stage 07 finds a design problem rather than a bug → back to 04 or 02

Going backwards is cheap. Shipping a wrong spec is not.

## What each gate is for

**Gate 1 — after the risk pass.** You answer the open questions. This is
where product decisions get made by a person. Do not let the agent choose
defaults here; the whole point of stages 01–03 was to surface these.

**Gate 2 — after the scaffold.** Types, signatures and failing tests exist;
no logic does. This is the cheapest possible moment to say "that is the
wrong shape". Read the scaffold properly — five minutes here saves an hour
of rework.

That was one of 19 files

The rest of Spec-to-Ship is written the same way. $49 once, yours permanently, 14-day refund.

# How to run the pipeline ## The shape Nine stages, two gates. The gates are the only places a human has to think hard; everything else is mechanical once the decisions are made. ## Context management Stages 01–05 can share one session. They are all about the same question (what are we building) and the accumulated understanding helps. **Start a fresh session at stage 06.** Paste in the spec and the plan. Nothing else. The exploratory conversation from stages 01–03 is now debris: it contains rejected options, earlier phrasings of the requirement, and half-formed ideas the model cannot distinguish from decisions. **Start another fresh session at stage 07.** Self-review inside the implementing context is close to worthless — the model has just spent an hour convincing itself the implementation is right. Give the review a clean context with the spec and the diff, and nothing about how the diff came to be. ## Time budget For a task that would take a competent developer a day: 01 Intake 2 min 02 Spec 10 min, mostly you reading 03 Risk pass 5 min GATE 1 however long the decisions take — this is the real work 04 Plan 5 min 05 Scaffold 10 min GATE 2 5 min reading 06 Implement the bulk 07 Self-review 10 min 08 Backfill 10 min 09 Writeup 3 min If stages 01–05 take longer than the implementation, either the task is genuinely ambiguous (good — you found that cheaply) or it is too small for the pipeline. ## Going backwards The pipeline is not a ratchet. - Stage 03 finds a risk that changes the requirement → back to 02 - Stage 05 cannot express the shape cleanly → the spec is wrong, back to 02 - Stage 06 hits something the plan did not anticipate → back to 04, not forward with an improvisation - Stage 07 finds a design problem rather than a bug → back to 04 or 02 Going backwards is cheap. Shipping a wrong spec is not. ## What each gate is for **Gate 1 — after the risk pass.** You answer the open questions. This is where product decisions get made by a person. Do not let the agent choose defaults here; the whole point of stages 01–03 was to surface these. **Gate 2 — after the scaffold.** Types, signatures and failing tests exist; no logic does. This is the cheapest possible moment to say "that is the wrong shape". Read the scaffold properly — five minutes here saves an hour of rework.