interview-me

interview-me

Onboarding + light user research in one flow. Get the basics right so downstream skills don't fumble identity; capture goals + current tools so the product team learns what users actually want.

flow

1. Pre-fill before asking. Run my_profile and get_person({ person_id: "me", include: "dossier" }) first. Use what comes back to:

   • Pre-select chips for Q1 (role) and Q2 (location) if there's a clean signal.
   • Draft Q3 (what you're building) if company context allows.
   • Collapse Q1 + Q2 into a single confirm if both are strong signal: "Here's what I have on you — Founder, NYC. Anything to correct?"

2. Calibrate framing to confidence. Confident and wrong is worse than guessed and humble.

   • Strong signal (clean my_profile field, unambiguous LinkedIn role/location) → present as confirm: "Founder at noticed, NYC — right?"
   • Thin signal (inferred from company domain, ambiguous title) → present as soft guess: "Looks like you're a founder? Tell me if I'm off."
   • Never present a thin inference as fact.

3. Ask one question at a time. Chips for MC; free text where chips don't fit. "Skip" is always allowed. The agent may rephrase the prompt copy to flow naturally, but chip option labels must match this spec exactly (consistency matters for research aggregation across users).

4. After the last question, show a quick summary of all answers and ask once: "save?"

5. On "save", fan out writes (see writes section). Echo what landed.

the questions

Ask in order. Pre-fill where signal exists. Skip Q1 / Q2 if covered by the collapsed confirm.

Q1. What's your role? (multi-select chips) Founder · Sales & Growth · Marketing & Brand · Partnerships & BizDev · Product & Engineering · Customer Success · Operations · Investing · Recruiting · Other "No big deal if you don't fit neatly. This is just to get started."

Q2. Where are you based? (single-select chips + "other" → text — never multi) NYC · SF · LA · London · Lisbon · Berlin · Remote · Other (type it)

Q3. What are you building or working on right now? (short free text) One line is fine.

Q4. Which of these are on your plate right now? (multi-select chips) Sales · Hiring · Find work · Fundraising · Investing · Customer success · Networking · Collabs · Other

Q5. If noticed nailed one thing for you, what would it be? (rank all 5 from most to least important)

• Remember context before a meeting
• Reconnect with people I've lost touch with
• Get warm intros to people I want to meet
• Find someone in my network who can help with something specific
• Remember names + context of people I just met

Use a drag-to-rank UI if available; otherwise ask the user to list the five in order.

Q6. How do you remember relationships today? (multi-select chips) Notes app · Spreadsheet · LinkedIn · Notion · CRM · Nothing · Other

Q7. Anything else worth knowing? (free text, skippable) Whatever doesn't fit above — who you're trying to meet, specific goals, context I should have.

pre-fill rules

• Q1 (role): match my_profile.role against the chip options. Strong signal → pre-select + confirm. Thin signal → soft-guess.
• Q2 (location): match my_profile.location to one of NYC / SF / LA / London / Lisbon / Berlin / Remote (case-insensitive; aliases ok: "New York" → NYC, "San Francisco" → SF). Confirm if clean; ask if ambiguous.
• Q3 (building): if my_profile or get_person('me').default_notes mentions a company/project, draft a one-liner and offer to edit: "Looks like you're building noticed — a personal networking agent. Refine?"
• Q4 – Q7: no pre-fill. Always ask.

If Q1 and Q2 are both strong signal, collapse them into one confirm shown before Q3. If the user corrects, fold corrections into the saved state before moving on.

writes (on save)

Identity → update_person('me', default_notes)

Read existing notes via get_person({ person_id: "me" }) first, then append a new dated block. Never overwrite.

<existing default_notes>

[mcp · skill:interview-me · YYYY-MM-DD]
Role: <Q1, comma-separated> · Based in <Q2> · Building: <Q3>

Research / goals → memory_save (4 entries, all prefixed for filtering)

• [mcp · skill:interview-me] active goals: <Q4 comma-separated> — category: commitment
• [mcp · skill:interview-me] noticed value ranking: 1) <top> · 2) <#2> · 3) <#3> · 4) <#4> · 5) <#5> — category: preference
• [mcp · skill:interview-me] current relationship-memory tools: <Q6 comma-separated> — category: fact
• [mcp · skill:interview-me] additional: <Q7> — category: fact (skip if blank)

memory_save dedupes server-side, so re-runs don't double-store.

Echo

After writes land, recap exactly what was saved: "saved your role + location to your record. saved your goals, value ranking, current tools, and notes to memory. you can re-run anytime to refresh."

the rule

• One question at a time. Always skippable.
• Chips for MC questions; chip option labels must match this spec exactly. Agent may rephrase prompt copy, never chip labels.
• Confident and wrong is worse than guessed and humble. Pre-fills surface as confirms or soft guesses, never as facts.
• Pre-fill + skip-if-known when there's signal; never invent answers when there isn't.

re-runnability

Idempotent. Re-running:

• Re-asks each question (with pre-fills from the latest profile state).
• Appends a new dated [mcp · skill:interview-me · YYYY-MM-DD] block to default_notes — older blocks stay as history, latest reflects current state.
• memory_save dedupes via embedding so research entries don't double-store.

retrieval note for downstream skills

Basic identity (role, location, what the user is building) lives in default_notes on the 'me' person — read via get_person({ person_id: "me" }). Goals, primary value ranking, current tools, and the open-ended add-on live in memory_search (filter on the [mcp · skill:interview-me] prefix). Both surfaces are needed for a complete picture of the user.

tool needs

• noticed: my_profile, get_person, update_person, memory_save
• No web search — V2 doesn't enrich beyond what my_profile already has.

explicitly NOT in scope

• Auto-enriching profile data beyond what my_profile already has.
• Branching question flows based on prior answers.
• Editing tags / relationship_types on 'me'.
• Anything that takes more than ~3 minutes.
• Welcome-first / parallel enrichment kickoff with "takes ~2 minutes" framing (V3 only).