interview-facilitation — independently scanned and version-tracked by SaferSkills.
SaferSkills independently audited interview-facilitation (Agent Skill) and scored it 96/100 (green). The audit ran 55 deterministic rules across Security, Supply Chain, Maintenance, Transparency, and Community; it found 0 high-severity and 1 lower-severity findings. The full rule-by-rule trace and per-finding evidence are below. Free, methodology-open.
Findings & checks · 1 flagged
The text {match} tells the agent to skip the normal "ask the user first" gate. Used adversarially it removes the human-in-the-loop check before destructive or sensitive actions, turning a normally-gated agent into a fire-and-forget executor.
Every scanned point with the score it earned and what moved between them.
First recorded scan — no prior version to compare against.
The primary manifest — the file an agent reads to learn what this artifact does.
A swarm interview is choreography, not free-for-all. Without structure, six personas talking at once is six monologues. With structure, it's a roundtable.
This skill covers the interview arc — how the swarm's questions are sequenced, how depth and breadth trade off, and how the user is given enough air to think between rounds.
Diverge first, converge last. The first round of questions should open the topic from many angles (each persona surfacing different concerns). The closing round should narrow to specific decisions. In between, the orchestrator decides whether one or two areas need deeper probing.
A well-facilitated swarm interview has three phases:
All personas spawn in parallel. Each surfaces their core questions, concerns, and observations. The user receives a synthesis covering breadth across all spawned personas.
This is the only round many brainstorms need. For a focused decision, opening + synthesis is sometimes enough.
If synthesis surfaced 1-2 areas of significant disagreement or critical open questions, spawn a focused second round with just the personas in tension on those specific issues.
Examples:
If the brainstorm produced a tentative direction, spawn a tight closing round to pressure-test the proposed decision. Usually 3-4 personas: Skeptic + relevant domain experts.
Closing prompt format: "Given that we're leaning toward [decision], each of you: name the strongest case AGAINST this and the strongest case FOR. One paragraph each."
The orchestrator constructs the prompts each persona receives. Question design matters even though the persona subagents have their own voice.
| Type | When to use | Example |
|---|---|---|
| Open | Phase 1 — divergent opening | "What's your perspective on adding offline mode?" |
| Probing | Phase 2 — targeted probing | "Skeptic, you said 'we're inventing demand.' Specifically — name 3 user requests in your imagined evidence base that would convince you the demand is real." |
| Closing | Phase 3 — convergent closing | "Given the direction is opt-in offline mode for paid users only, name the strongest argument against." |
⛔ Avoid leading questions in Phase 1. They constrain the persona's contribution.
| Leading (avoid in Phase 1) | Neutral (preferred) |
|---|---|
| "Why is offline mode important?" | "What's your perspective on offline mode?" |
| "How should we ship this default-on?" | "How should we approach the rollout?" |
| "What are the risks of NOT doing this?" | "What are the relevant risks?" |
In Phase 2 and 3, leading questions become useful — they probe specific areas.
A good prompt covers:
A prompt that asks only "How should we do X?" misses the "should we" question. A prompt that asks only "Why X?" misses the implementation reality.
The persona's structured output formats handle the balance — but the orchestrator's framing should leave space for all three.
The fundamental tradeoff in interview design:
| Approach | Pros | Cons |
|---|---|---|
| Breadth (many personas, one round) | Surfaces diverse perspectives quickly | Each contribution is shallow |
| Depth (few personas, multiple rounds) | Pursues specific tensions | Misses perspectives not in the few |
The standard arc balances them: Phase 1 is breadth (6 personas, one round), Phase 2 is depth (2-3 personas, focused probing), Phase 3 is breadth-on-narrow-decision.
See references/interview-arcs.md for the full library of arc patterns.
Don't auto-spawn the next round. Between Phase 1 and Phase 2, present the synthesis to the user and ask:
The user might be ready to decide after Phase 1. Don't keep spawning if they're done.
Standard topic + context template. Personas know what to do.
Topic: [one-sentence framing]
Context:
- [Background]
- [Constraints]
- [Specific question being decided]
Your job: Contribute your perspective in your characteristic voice.Tighter, calls out the specific tension.
Specific tension from Round 1:
- [Persona A] said [position]
- [Persona B] said [opposing position]
Your job (you are [Persona A or B]):
- Defend your position with specifics
- Name the conditions under which the other side would be right
- Identify the evidence that would settle this disagreementFixed format around the proposed decision.
We're leaning toward [decision].
Your job:
- Strongest argument FOR (one paragraph)
- Strongest argument AGAINST (one paragraph)
- The decisive consideration in your view (one sentence)See references/question-design.md for detailed prompt patterns and worked examples.
swarm-protocolswarm-synthesiscustom-personasdeep-interview (skillstack)elicitation (skillstack)What it looks like: spawning round after round, going from 6 personas to 12 to 18 in pursuit of "completeness."
Why it's wrong: synthesis becomes harder; the user can't act on infinite input.
What to do instead: most decisions are well-served by 1-2 rounds. Push for a decision after the second round. If the user genuinely needs more, you have a decomposable topic — split it.
What it looks like: Phase 1 ends, orchestrator immediately spawns Phase 2.
Why it's wrong: the user might have wanted to think, or might be ready to decide. Burning compute on rounds the user doesn't need.
What to do instead: present synthesis after each phase, ask whether to continue.
What it looks like: asking "Skeptic, what could go wrong with this great proposal?"
Why it's wrong: baked-in framing ("great proposal") biases the persona; defeats the point of the skeptic.
What to do instead: "Skeptic, your perspective on this proposal." Let the persona apply their voice without orchestrator framing.
What it looks like: Phase 3 prompt that pressures personas to converge ("everyone, agree on the right answer").
Why it's wrong: the swarm's value is preserved dissent. Forcing consensus loses the disagreement.
What to do instead: Phase 3 prompts ask each persona to identify the "decisive consideration in their view," not to align with others.
What it looks like: every persona's prompt includes 1500 words of context, prior synthesis, history.
Why it's wrong: personas are fresh subagents; bloat dilutes the prompt; they can't synthesize from a wall of text.
What to do instead: keep persona prompts to 100-300 words of context. The persona's own contribution-shape handles the rest.
The standard arc:
| File | Contents |
|---|---|
references/interview-arcs.md | The 4 standard arcs (divergent-only, probing, full-arc, decision-pressure-test); when each works |
references/question-design.md | Prompt patterns for each phase; worked examples; leading-vs-neutral library |
references/depth-vs-breadth.md | When to favor each; how the topic shapes the choice |
~30 seconds. Free. No account. Every finding cites a rule and a line of evidence.