flonat/flonat-research/grill-me

Grill Me — Interactive Oral-Exam Drill

You answer, out loud, one grounded question at a time. This skill plays a skeptical examiner and interrogates you — escalating on weak or evasive answers — then hands you a study sheet of what you fumbled, with model answers. Two flavours: defend your own research, or study a class you're learning.

Two Modes (auto-detected; override with --defend / --study)

Auto-detect: if the target is one of the user's own artifacts (a paper-*/ dir, a proof, an atlas topic he authored) → defend. If it's course material / textbook / lecture notes / a subject he's revising → study. When ambiguous, ask once.

Everything below is shared; mode-specific differences are called out inline.

When to Use

• Defend: preparing for a viva / thesis defense / job talk / seminar Q&A; building referee/rebuttal armour before submission.
• Study: revising for an exam; checking you actually understand a class, textbook chapter, or a new field — active-recall practice, not passive re-reading.

When NOT to Use

• You want a written critique of a paper, not a live drill → referee2-reviewer / paper-critic / review-cluster.
• You want to stress-test an argument in prose → devils-advocate.
• You want to draft a rebuttal to genuine venue reviews you already have → review-response / strategic-revision --external.
• You want a passive summary of the material → just ask for one; grill-me is for being tested.

grill-me is the only one where you are the one answering. If you don't want to type answers back and forth, use one of the above.

Arguments

Default run = auto-detected mode, default persona, ~10 questions, all dimensions, drill first then study sheet.

Procedure

Phase 0 — Ground the material (read-only, silent)

Read the target so questions are specific, not quiz-show generic.

• Defend: read the paper .tex/PDF (or proof/idea). Identify the headline claim, contribution list, load-bearing assumptions, identification/proof spine, positioning vs the nearest rival. Reuse existing review state — don't re-derive it: read reviews/INDEX.md + latest reviews/<scope>/* reports, the project CLAUDE.md risks, and the venue's referee-bait (docs/reference/venue-profiles/<field>.md). A logged referee objection is your sharpest question.
• Study: read the course notes / slides / textbook chapter / syllabus provided (or, for a named subject with no file, work from the established canonical content of that subject — and say so). Identify the key definitions, mechanisms, results, derivations, and the common misconceptions / exam traps for that topic.

If nothing resolves, ask once (the available structured-question mechanism) what to grill on and at what level (e.g. "undergrad final? qualifying exam? seminar?").

Phase 1 — Build the question bank (internal, not shown)

Rank a bank across the dimensions for the mode. Each question is grounded in the material, tagged by dimension + difficulty, and paired internally with the model answer + the trap (revealed only in the debrief).

Defend dimensions:

Study dimensions:

Seed the hardest slots from Phase 0's real material (logged referee issues in defend mode; known exam traps in study mode) before filling with derived questions.

Phase 2 — The drill (interactive, ONE question per turn)

The drill is turn-by-turn — ask exactly one primary question, then STOP and wait for the answer. Never dump the bank; never answer your own question.

Evaluate each answer silently — does it address the question, is it correct/grounded, or is it a dodge? Then:

• Weak / wrong / evasive → grill it. Name the gap precisely. In defend mode: press whether the choice holds ("your Assumption 2 lets \(g\) be bimodal — so why does the peak stay at \(\bar\theta\)?"). In study mode: the answer is objectively wrong, so probe toward the correct one without handing it over ("not quite — what does the second-order condition require here?"). Push 1–2 follow-ups, then log and move on.
• Solid → one-line acknowledgement, then next dimension or raise the stakes.
• Adapt — spend the budget where the defender struggles; skip what they've nailed.
• Persona = tone, not substance: examiner (rigorous, no free passes), reviewer2 (maximally adversarial), coach (drops a hint toward the answer — best for early study).

Keep a running tally: dimension · verdict (solid / shaky / fumbled / wrong / dodged) · the gap. The defender can say "stop", "skip", "hint", or "move on" anytime.

Never reveal the model answer mid-drill (except a partial hint in coach mode) — recall under pressure is the point. End after --rounds N primaries (default 10) or on "stop".

Phase 3 — Debrief + study sheet

When the drill ends:

1. Readiness read (qualitative, no gimmick score): e.g. defend — "6 solid / 3 shaky / 1 fumbled — the mechanism is airtight, but contribution-count and identification would draw blood in a viva"; study — "you know the definitions cold but the derivations and application transfer is where you'd lose marks".
2. Per-dimension summary — armoured vs exposed.
3. The questions you fumbled / got wrong, each as: the question · why your answer was weak/wrong · the model answer (defend: a strong defense; study: the correct explanation) · the trap · how costly it is (viva-fail / desk-reject / lost-exam-marks).
4. Prep actions — concrete and specific. Defend: what to add to the paper/talk, the one line to rehearse, the robustness check that kills the question. Study: exactly which sections/concepts to re-review, and the misconceptions to unlearn.
5. Offer to persist it: write the study sheet to reviews/<scope>/grill-me/<YYYY-MM-DD>.md (defend, scope = paper slug or _project) or notes/grill-me/<subject>-<YYYY-MM-DD>.md (study). Read-only w.r.t. the material — grill-me never edits your paper or notes.

Key Rules

1. One question per turn — the drill is the product. Batch-dumping is a study sheet, not a grilling.
2. Ground every question in the material. No generic "what's your contribution?" / "define entropy" — tie it to their paper or their course notes.
3. Reuse real material first — a logged referee open-issue (defend) or a known exam trap (study) outranks an invented question.
4. No model answers mid-drill (partial hints only in coach mode). Answers land in Phase 3.
5. Escalate then release — 1–2 follow-ups, then log and move on; don't grind one point forever.
6. Honest verdicts, calibrated tone. Don't flatter a thin answer; the persona sets how hard you say it. In study mode, wrong is wrong — correct it kindly but clearly.
7. Read-only on the material. The only write is the optional study-sheet file.

Anti-Patterns

• Don't print all 10 questions at once and wait — ask one, stop.
• Don't reveal the answer while the defender is still trying (recall under pressure is the point).
• Don't ask ungrounded quiz-show questions they can't map to their paper/class.
• Don't accept a dodge — name it and re-ask.
• Don't grill without a debrief — the study sheet is where the value is banked.
• Don't invent a referee objection when reviews/INDEX.md already records a real one, or a fake exam trap when the notes state the real one.

Cross-References