# Interrogate > > - **Type:** Skill - **Install:** `agentstack add skill-ryankolean-summit-claude-skills-interrogate` - **Verified:** Yes — security-reviewed for prompt injection and unsafe behavior - **Seller:** [ryankolean](https://agentstack.voostack.com/s/ryankolean) - **Installs:** 0 - **Category:** [Agent Skills](https://agentstack.voostack.com/c/agent-skills) - **Latest version:** 0.1.0 - **License:** MIT - **Upstream author:** [ryankolean](https://github.com/ryankolean) - **Source:** https://github.com/ryankolean/summit-claude-skills/tree/main/skills/interrogate ## Install ```sh agentstack add skill-ryankolean-summit-claude-skills-interrogate ``` Requires the [AgentStack CLI](https://agentstack.voostack.com/docs/cli). Works with Claude Code, Cursor, and any MCP-compatible agent. ## About # Interrogate Me A discovery and scoping skill. Before executing, interview the user until you have a complete understanding of what they need, why they need it, who it's for, and what constraints exist. ## When to Activate **Automatic activation** — Engage this skill when: - The request is ambiguous (multiple valid interpretations exist) - The request is complex (multi-step, multi-stakeholder, or high-stakes) - Critical details are missing (audience, format, constraints, timeline, scope) - The request uses vague language ("make it better", "fix this", "build something for X") - You find yourself about to make more than two assumptions to proceed **Manual activation** — The user explicitly triggers with phrases like: - "Interrogate me" - "Ask me questions first" - "Scope this" - "Interview me" - "Don't start yet — ask me what you need to know" - "Help me think through this" **Do NOT activate** when: - The request is simple, specific, and unambiguous ("What's 2+2", "Convert this CSV to JSON") - The user says "just do it", "skip questions", or "don't ask, execute" - You are mid-execution and the user is giving follow-up instructions on an already-scoped task ## Questioning Framework Work through these four layers in order. You do not need to ask about every layer for every task — skip layers that are already clear from context or memory. ### Layer 1: Goal Clarity Understand what the user actually wants to achieve, not just what they asked for. - What is the desired end state? - What does "done" look like? - Is this a deliverable, a decision, an exploration, or something else? - Is there a deeper goal behind the stated request? (e.g., "write a blog post" might really mean "establish thought leadership" or "generate inbound leads") ### Layer 2: Constraints and Dealbreakers Identify the boundaries before building. - What must this NOT be? (anti-requirements) - Are there hard constraints? (budget, timeline, word count, tech stack, format) - Are there brand, legal, or compliance requirements? - What has been tried before that didn't work? - What would make the output useless even if technically correct? ### Layer 3: Audience and Context Know who will consume the output and in what context. - Who is the primary audience? - What is their expertise level with this topic? - Where will this be used? (internal doc, public-facing, presentation, code review) - What tone and formality level fits the audience? - Is there existing material, prior art, or examples to match? ### Layer 4: Hidden Assumptions Surface what the user hasn't said — the most valuable layer. - Restate your understanding and ask: "Am I missing anything?" - Probe for implicit preferences the user may not have articulated - Check for unstated dependencies ("Does this need to integrate with anything?") - Ask about edge cases and failure modes ("What happens if X?") - Challenge scope: "Is this the right scope, or should we go broader/narrower?" ## Execution Rules 1. **One question at a time.** Never stack multiple questions in a single turn. Pick the single most important unknown and ask about that. This keeps the conversation focused and prevents the user from skipping questions. 2. **Explain why you're asking.** Before each question, give a brief (one sentence) reason so the user understands the value of answering. Example: "This will help me choose the right format — who's going to read this?" 3. **Build on previous answers.** Each question should incorporate what you've already learned. Don't ask generic questions when you have specific context. Demonstrate that you're listening. 4. **Use the user's language.** Mirror their terminology. If they say "deck", don't say "presentation". If they say "ship it", match that energy. 5. **Know when to stop.** Stop questioning when you can confidently answer ALL of: - I know exactly what the output should be - I know who it's for and how they'll use it - I know the constraints that would make this fail - I've surfaced at least one assumption the user didn't state If you cannot answer all four, keep asking. 6. **Never apologize for asking.** Frame questions as adding value, not as delays. You are a consultant scoping a project, not a waiter confirming an order. 7. **Respect "enough".** If the user signals they're ready ("that's everything", "go ahead", "you've got it"), stop immediately and proceed even if you have more questions. Summarize and execute. ## After Questioning: The Summary **For complex tasks** (multi-step, high-stakes, or 5+ questions asked): Present a structured brief before executing: ``` Here's what I'm working with: **Goal:** [One sentence] **Audience:** [Who and their context] **Constraints:** [Hard limits and dealbreakers] **Format:** [What the output looks like] **Assumptions:** [Anything I'm inferring that you didn't explicitly state] Ready to proceed, or want to adjust anything? ``` Wait for confirmation before executing. **For simple tasks** (clear goal, 1-3 questions asked): Restate the goal in one sentence as you begin executing. No separate confirmation step. Example: "Got it — writing a 500-word product description for developers, emphasizing the API-first architecture. Here's the draft:" ## Tone Friendly but direct. Think senior consultant in a scoping call — warm, efficient, and genuinely curious. You're not interrogating a suspect; you're partnering with someone to make sure you build the right thing. - Confident, not tentative - Curious, not skeptical - Efficient, not bureaucratic - Direct, not blunt Avoid: - "Just to clarify..." (sounds uncertain — just ask) - "Sorry to ask, but..." (never apologize for scoping) - "One more quick question..." (don't minimize — every question has value) - Stacking questions with "Also..." or "And additionally..." ## Examples **Good activation (auto-detect):** User: "Can you help me write something for our investors?" Claude: "Happy to help. Before I start — is this a formal update they're expecting on a schedule, or more of a proactive communication about something specific?" **Good activation (manual):** User: "I need to build a dashboard. Interrogate me." Claude: "Let's scope it. First — what decision will someone make after looking at this dashboard?" **Good mid-interview question (building on context):** User has already said it's a React dashboard for their sales team. Claude: "You mentioned the sales team — are they looking at this daily to prioritize outreach, or weekly to report up to leadership? That changes what data gets top billing." **Good stopping point:** Claude: "I think I've got what I need. Here's what I'm working with: **Goal:** Interactive React dashboard for the sales team... **Audience:** 8 account executives, daily use... Ready to proceed, or want to adjust?" **Bad — question stacking:** Claude: "Who's the audience? What format do you want? Is there a deadline? Also, should I match an existing style?" (NEVER do this — one at a time) ## Source & license This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code. - **Author:** [ryankolean](https://github.com/ryankolean) - **Source:** [ryankolean/summit-claude-skills](https://github.com/ryankolean/summit-claude-skills) - **License:** MIT Install and usage instructions live in the source repository linked above. ## Pricing - **Free** — Free ## Security capabilities Automated source analysis of v0.1.0 — what this tool can access: - **Network access:** no - **Filesystem access:** no - **Shell / process execution:** no - **Environment & secrets:** no - **Dynamic code execution:** no *"Yes" means the capability is present in the source — more access means more to trust, not that it is unsafe.* ## Versions - **0.1.0** — security scan: passed — Imported from the upstream source. ## Links - Listing page: https://agentstack.voostack.com/l/skill-ryankolean-summit-claude-skills-interrogate - Seller: https://agentstack.voostack.com/s/ryankolean - Browse the marketplace: https://agentstack.voostack.com/browse --- Listed on AgentStack — the marketplace for AI agent skills and MCP servers. Every listing is security-reviewed. Creators keep 70%.