eg-interview

/eg-interview — interview-strategy intake (grill-me lineage)

The top-of-funnel discipline for the human-facing agent: before decomposing or delegating a request, extract what the requester actually needs — by asking, not by assuming. The output is a delegable brief that travels with every downstream delegation, so worker agents execute a confirmed intent instead of an upstream guess.

1. When to run (and when not to)

Run when a request is broad, ambiguous, or high-stakes and will fan out — a program-scale ask, a "make X better" with no criteria, work that commits resources or publishes externally. Do NOT run for requests that are already specific and cheap to redo: interviewing a clear request is over-questioning, which costs requester patience for zero information gain. Rule of thumb: if you can state the purpose, the deliverable, and the success check in one sentence each and would bet on being right, skip the interview and confirm inline instead.

2. What to press for (the six extraction targets)

1. Purpose — why this, why now; what changes for the requester when it lands. The stated task is often a means; the purpose is what survives redesign.
2. Deliverable shape — what artifact/state counts as the output (a document? a running service? a decision?).
3. Constraints — deadline, budget/effort ceiling, technology or policy boundaries, do-not-touch zones.
4. Success criteria — how the requester will judge it done; what verification they'd accept.
5. Scope edges — what is explicitly in and out; the adjacent work they do not want started.
6. Priority trade-offs — when speed, completeness, and cost collide, which yields first.

3. Question discipline

• Batch few, high-yield — 2–4 questions per round, at most ~2 rounds. Each question must change what you'd do next; drop any whose every answer leads to the same plan.
• Default-and-confirm over open-ended — where an answer is inferable, state the inferred default and ask for correction ("기본값 X 로 진행할게요 — 아니면 알려주세요") instead of asking from zero.
• Structured choices where the host supports them — present options with costs/benefits and a recommendation. Compose with Hyperbrief for the presentation register: default the option prose to the plain-language levels (L1.1.1–L1.2.2) and honor an "explain more simply" fallback — the requester picking an option must actually understand it.
• Stop on diminishing returns — when the remaining unknowns are cheaper to resolve by doing (a reversible probe) than by asking, stop interviewing and mark them as assumptions.