Figure It Out

---
name: figure-it-out
description: "Design and execute an auditable plan for a complex task that lacks a narrower workflow."
disable-model-invocation: true
---

# Figure it out

Use when no narrower workflow owns the requested outcome. First state a
falsifiable done condition, scope, estimated units, dependencies, and what
would stop the run. Choose verification rigor from reversibility, blast radius,
and cost of a wrong result. For extended work, show the phase plan before a
long run so the user can redirect; authorized reversible work can continue.

Design independently verifiable units. Put the riskiest unknown and its proof
early; capture the baseline before changes. Use [architect](../architect/SKILL.md)
for a consequential one-way design, not for mechanical work. Parallelize only
independent ownership, with isolated writers. Name the artifact and check for
each unit, along with any publication or human decision gate.

Run a hypothesis loop for each unit: state prediction, make one bounded change,
measure on the relevant surface, inspect the artifact, and decide VERIFIED,
NOT VERIFIED, or INCONCLUSIVE. A child report is not root proof. Reject or
revise an unsuccessful change without discarding unrelated user edits. If the
measurement gate is wrong, correct it explicitly before relying on new runs.
Record decisions as work proceeds with [show-me-your-work](../show-me-your-work/SKILL.md).

At the end, check the complete outcome against the original predicate on the
real artifact. Encode repeated mistakes in a test, schema, or check when that
is warranted. Return the designed phases, decision trail, per-unit verdicts,
overall proof, and open gaps. This workflow does not imply unbounded autonomy,
an unattended schedule, or authority to publish.