reverse-engineer-testslisted
Install: claude install-skill sub7an/reverse-engineer-tests
# Reverse Engineer Any Test (RET)
A test paper leaks its own construction. Every question is built to a
template: one tested skill, a planted trap, and wrong choices picked to catch
predictable mistakes. Recover the templates from real past papers, count
them, and check each claim against papers it was not learned from.
The skill's value is honesty. Most test-prep "tricks" fail when counted, and
a student is better served by "this trick is a myth, here are the numbers"
than by a confident rule that costs marks. Report busted myths as
prominently as real signals.
## Two rules that hold everywhere
1. **Evidence, not vibes.** A pattern is real only when seen in 2+ questions
with cited question ids (qids). One occurrence is an anecdote. Never
estimate a frequency; the scripts count. Keep coverage separate from
completeness: a pattern absent from the papers analyzed is not proven
absent from the test. `references/evidence-model.md` defines the record
format, unknowns and the mandatory footer; read it before Phase 2.
2. **No signal without its held-out value.** A p-value says a pattern
exists; it does not say it earns marks. Phase 5 prices every
answer-choice signal on papers it never saw. If that gain's interval
includes 0, say "not established".
## Pick the lane first
Run only the phases the request needs.
| The user wants | Run |
|---|---|
| "Crack this test" with whole papers | Phases 1–5, then 7–8 if asked |
| "Which topics / patterns matter" | P