Calibrate — size the work before structuring it

One judgment, made once, scales everything downstream. The cost of misjudging is asymmetric: too many subdivisions create coordination overhead, duplicated context, and drift between documents; too few create units an agent cannot finish in one clean context. So spend real thought here, then commit.

Guiding principle: minimize the number of subdivisions — without sacrificing quality, clarity, or performance. When two structures both work, pick the one with fewer pieces.

What to weigh (and what to ignore)

Assess exactly these four dimensions. Do not use file count or lines of code — a 40-file mechanical rename is lighter than a 3-file novel algorithm.

1. Customer-facing feature count — how many distinct things a user will see or do. The primary driver for initiative splits.
2. Coupling — do the features share a schema, a UI surface, a flow? Tightly-coupled features belong in one unit (splitting them creates ordering pain); separable features are split candidates.
3. One-shot executability — could a single agent execute this unit start-to-finish without exhausting or rotting its context? Think: how much must be held in memory at once (modules touched, contracts to respect, states to track). If the honest answer is "it would be drowning by the end", the unit is too big.
4. Requirement clarity — vague or contested requirements need more questions and smaller, safer units; crisp requirements support bigger ones.

Focus modes

The caller passes a focus. Return the matching number(s):

Reference points (from the OpenLatch product vault): Admin Broadcast Messaging (one feature surface, deeply coupled internals) → 0 initiatives. Agent Scan PRD (engine + report UI + website refresh — three separable, independently-shippable clusters) → 3 initiatives (I-5.5/5.6/5.7).

Output — the verdict block

Always return exactly this, so calling workflows can consume it directly:

CALIBRATE VERDICT
Tier: Light | Medium | Heavy
<Focus numbers — e.g. "Initiatives: 2 (engine, report-ui)" or "Plans: 3" or "Question budget: 20">
Why: <2–3 lines citing the drivers: feature count, coupling, one-shot executability, clarity>

State the verdict out loud so the user can correct a misread early. If later steps (clarify-intent answers, research findings) materially shift the scope, the caller should re-run calibrate — say so when the risk is visible.