Skill
A four-move design discussion the model drives, the user rules.
/attune:explore Explore runs a structured design discussion through Ask, Sort, Research, and Experiment. The model investigates purpose, generates real alternatives, hunts for shortcuts, and settles world facts through the cheapest sufficient source — bringing the user only their own facts and rulings, so decisions rest on evidence and taste, never assumptions.
Ask raises purpose, alternative, and shortcut questions aimed at the model itself. Sort classifies each as world fact, user-held fact, or judgment. Research settles world facts through memory, machine, then web. Experiment builds the answer nothing existing holds.
The model does the legwork — climbing the request's five levels, generating alternatives that differ in mechanism, and seeking shorter paths — so the user spends judgment on taste, priorities, and risk, never on research they could have done themselves.
Most questions end at memory or this machine without touching the web. Volatile facts are never treated as stable, load-bearing recalls are verified before the recommendation rests, and the web is gated behind a single scouting pass per settled intent.
Web researchers write dated briefs to disk as the evidence trail; the thread carries only the compiled digest. Conflicts are resolved by authority and recency, unopened sources are marked unverified, and trials run against the real artifact — never a stub from memory.
Raise purpose, alternatives, and shortcut questions aimed at the model.
Investigate what the user is ultimately trying to make happen, climb the request's five levels, generate alternatives that differ in mechanism, and seek a shorter path — adding every question to a visible open list.
Classify each open question once, before any work proceeds.
Mark each question as a world fact, a user-held fact, or a judgment; route user-held facts and judgments to the user, and decompose compound questions so each part is routed on its own.
Settle world facts through the cheapest sufficient source.
Move through memory, this machine, then the web — escalating only when a source cannot settle the question — and verify each load-bearing recalled fact before the pass presents its recommendation.
Build the answer nothing existing holds.
When research finds nothing or sources conflict on an observable claim, run the smallest rig that yields the observation against the real artifact, recording every trial's command and output as created evidence.
Deliver the pass's understanding, alternatives, and recommendation.
State purpose and intent, the alternatives explored, the strongest shortcut, the recommendation with rationale, and what remains unverified, assumed, or parked — each labeled as what it is.