Use when a meta-agent request is too ambiguous to safely generate, package, publish, or adapt without one to five targeted questions.
SkillScore73Maintained
Automatic
◆AI Engineer pack
Last verified 8/17/2026
What it changes
Passed 15 of 21 test cases with the skill, and 9 of 21 without it.
Without skill41%
With skill73%
Checks it passed that the model alone missed
Re-runs mode classification.
Asks between 1 and 5 questions.
Asks an initial interview batch of 8 to 12 questions.
and 6 more in the full report
7 fixed9 passed either way6 still failing0 broke
Pass-rate lift
+32 pp
Tokens
+27%
Agent turns
0%
Response time
-27%
Model gemini-3.6-flashMeasured 8/17/2026v216 of 22 cases passed with skill · 21 comparable
Tokens, turns and time describe resource use, not a dollar price.
See it in action
One of 6 test cases the skill fixed. Same model, same prompt, with and without the skill.
1 of 6
Prompt
The user just answered our batch of clarifying questions for a customer support agent package, revealing that handling complaints requires independent legal compliance approval. What is the immediate next step in the procedure before generating package files?
What was checked, without and with the skill
Checked
Without skill
With skill
Re-runs mode classification.
Missed
Passed
case-21 · gemini-3.6-flash · 8/17/2026 · answers are capped and redacted
What's inside
The instructions your model reads when the skill loads.
SKILL.md69 linessingle file
Clarify Question Loop
Ask only questions that change the generated package, runtime adapter, safety
boundary, or public/private release decision.
For /hep-build creation or behavior-changing packaging, this is not a
substitute for the Builder Interview and Research Gate in
contracts/builder-interview-research-gate.md. Run that gate first: ask an 8-12
question first batch, research similar agent repositories or comparables and
academic/professional theory, then use this clarify loop only for the remaining
narrow ambiguities.
Procedure
Classify the current best mode.
If single-agent vs team selection would change the package shape and the
independent ownership boundaries are unclear, ask before generation. The
first batch must include this plain-language question: "이 일을 한 명의
전문가가 처음부터 끝까지 맡으면 되나요, 아니면 조사/분석/검토처럼 여러
전문가가 나눠 맡고 마지막에 합쳐야 하나요?"
Follow up on role count, role-specific tools/permissions, whether outputs
must be synthesized, and whether artifacts are sequential dependencies or
independent parallel packets.
Identify missing facts that would change files or safety.
Ask one to five short questions, preferably three. If more than five
functional-quality questions remain, return to the Builder Interview and
Research Gate instead of pretending the package is ready.
Do not ask for secrets. Ask for secret names or setup boundaries instead.
After answers arrive, re-run mode classification if needed.
Generate or repair the package using the answers and list assumptions.
Budgets and stop rule (briefing interview engine)
This loop shares the briefing interview engine's contract
(agentlas_cloud/interview/): a question is only worth asking if the answer
would change execution, not just its phrasing. Respect the surface budget
(chat 3-5 in one batch, stormbreaker <= 8 across two batches, build 8-12 plus
follow-ups). 'decide later' is always a valid answer — record it as deferred,
never re-ask. When answers you auto-confirmed from code/memory reach three in a
row, the next question must go to the human.