Goal: given a chosen solution, map the solutions not taken — explore the option space horizontally and recommend a pick.

Companion to oracle (which explores one stack vertically — depth, unknown-unknowns). compass explores breadth across competing approaches.

When not to use: for open exploration — "how should I think about X", general research — the bare harness is the better default; it plans its own approach and reads more cohesively than a fixed procedure. Reach for compass when a choice is already made or leaning, and you want it attacked and a verdict issued.

Steps

Misfit check. compass misfits when no solution has been proposed — there is no baseline to compare against, so the skill would have to invent one, which is where it drifts generic. That is the whole test.

On misfit, name the absent trigger in one line, then answer directly and skip the rest:

Skipping compass: no solution proposed — no baseline to compare alternatives against. Answering directly.
Restate the proposed solution + intent. One line naming the chosen approach and the problem it solves. This is the baseline everything else is compared against.
Frame the axes that matter. Name the dimensions alternatives differ on for this problem — e.g. latency, throughput, ops cost, consistency, lock-in, team familiarity, scaling ceiling. These axes pick the field: the alternatives worth surfacing are the ones that win or lose meaningfully on them.

Aim: the axes dominating this decision are named, and every later alternative can be placed on them.