Goal: surface what the user does not know they don't know — the gaps beneath the prompt ("unknown unknowns", Rumsfeld) — and teach each one deep enough to stick.

Companion to compass (breadth across competing approaches). oracle descends one stack vertically.

When not to use: for open exploration — "how do I", "what is", general research — the bare harness is the better default; it plans its own approach and reads more cohesively than a fixed procedure. Reach for oracle when committing to a design in an unfamiliar domain and the risk is a gap you cannot name.

Steps

Misfit check. oracle misfits when the prompt is a factual lookup with a determinate answer — "what is X", "what's the syntax for Y", "does Z support W". Nothing is being committed to, so there is nothing to descend toward. That is the whole test; apparent simplicity is not a misfit, since a short prompt preceding a real decision is exactly where gaps hide.

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

Skipping oracle: factual lookup with a determinate answer — no pending design commitment to descend toward. Answering directly.
Restate + infer intent. Restate the prompt and state its underlying intention in one line.
Calibrate depth, not scope. Read the prompt's vocabulary for the level the user works at, and let it set how much explanation each gap gets — never which gaps you reach. Fluency at one layer says nothing about the layers beneath, so calibration must not prune step 4's descent. Descend fully; explain briefly where the user is plainly fluent.
Ground. Search before mapping any gap: WebSearch for what practitioners actually do; Context7 MCP for library/framework/API documentation. Rank sources by references/source-priority.md — read it before searching. Read each source for what it presupposes, what it warns about, and what it cites — that is the raw material step 4 descends from.