Harden

Travel six months forward, live with the target as its future users will, record what went wrong and what went right, then return to the present and rank the changes that would have prevented the incidents without killing the successes.

The target is whatever the user brings: a git diff, a source module, a spec or PRD, an API surface, a document, a skill, or an idea described only in this session. Adapt the simulation to what the target actually is — a diff has callers and edge cases, a spec has readers and ambiguities, a skill has invoking agents and misfires.

Brief

Establish two things before simulating: what the target is and who touches it downstream.

Explore the surrounding project for what the supplied context does not say — callers, adjacent modules, existing conventions the target must live beside. If the target is an idea with nothing in the project to explore yet, skip that exploration and ask the user directly about intended users, reach, and dependents. If the supplied context and the project together answer who uses this, how it is reached, and what depends on it, go straight to Timeshift. Ask only about the gaps that remain, and ask them in one batch.

Do not ask the user to predict the future. That is the skill's job.

Choose the simulation mode from the invocation: --mode subagents dispatches an independent subagent per lens, anything else runs inline. Default inline. Subagents cost more but stop one agent's blind spots from flattening all seven lenses into the same three obvious failures — use them on consequential or unusually broad targets.

Timeshift

Move to six months from today and write field reports — incidents and successes recorded as they happened, in the past tense, dated by month offset.