3.2 KiB
The planning dial — how much rigor a plan gets, chosen per piece
Planning is a spectrum, not a binary. When the wave model is selected, build care varies by wave composition; planning needs the same flexibility, or every piece gets one protocol — over-working the trivial and under-working the hard. Choose the rung from the work's uncertainty and consequences.
The rungs — dial up by what the piece is
- P0 — mini-requirements + pointers. A few observable outcomes and where to look; the builder figures out the how. Right for simple, reversible, well-trodden work.
- P1 — authored spec. Intent, mini-requirements, a failing-test-first proof contract, declared territory. The default.
- P2 — plus a research round, run BEFORE the spec freezes, never during the build. The research prompt is authored by an agent who holds the product context (see the gate below); execution may fan out to any agents — web legs, documentation legs, code-reading legs in parallel. Research returns as CONTEXT with citations, never prescriptions; synthesis against your own architecture and intent follows, and it differs for greenfield (design the thing) versus brownfield (sharpen your approach).
- P3 — plus an adversarial pass. A non-author who holds the product context attacks the plan in both directions with the product goals as judge. Amendments land with proof-contract teeth — a new or changed checkbox — never as prose advice.
- P4 — plus a blind design gate. A from-scratch design is committed before reading the priors, then diffed. Similar → proceed; significantly different → stop for a design session with the owner.
The context gate
Only agents with the product's world context installed may author research prompts, run adversarial passes, or weigh in on architecture questions. Execution researchers and builders need no such install — the agent CREATING their instructions must have it. The reason: research shaped by someone who does not know what the product is for returns answers to the wrong question, fluently.
What earns P2 or P3 — importance is not the test
A piece earns P2 when the spec would otherwise freeze on an unresolved external unknown — a feasibility door or an open design space only research can close. A piece earns P3 when its failure mode is invisible to its own author: load-bearing substrate where a plausible, self-consistent plan would ship the exact disease it exists to kill. The discriminator is "would a wrong plan be expensive, hard to undo, and invisible from inside" — never "is this piece central." Importance without an unclosable unknown or an author-blind failure mode stays at P1. Planning is text-only work; keep its scope proportional to the unresolved question and the cost of a wrong plan, rather than assuming a fixed duration for a rung.
Cost calibration
Scope research to the unknown that could change the plan; scope an adversarial pass to the failure modes the author may miss. Record the questions resolved and the resulting design changes so the owner can judge the value of the selected rung. Run planned research before build dispatch, while its answers can still shape the spec.