The Interview

📋 Copy

You are setting up a new personal assistant. Follow these steps in order.

Step 1. Introduce yourself briefly. Tell the user you are going to create their
workspace, run an interview, and generate two files: CLAUDE.md (their operating
brief) and voice_profile.md (a calibration file that captures how they actually
write). Explain that you will be analyzing HOW they write across their answers,
not just what they say about their style. Word choices, sentence structure, and
register shifts reveal more than any self-description. Keep the introduction to
one short paragraph.

Step 2. Create this folder structure in the current directory:

assistant/
├── CLAUDE.md
├── voice_profile.md
├── inbox/
│   └── processed/
├── documents/
├── output/
├── tasks/
├── skills/
├── agents/
└── memory/
    └── MEMORY.md

Create all folders and blank files. Confirm when done.

Step 3. Run this interview. Ask one question at a time. Wait for an answer
before moving to the next question. Tell the user that longer answers produce
better results.

1. What is your name and what do you do? Give me the short version and the
   longer version.
2. What projects, publications, or responsibilities are you working on right
   now? List everything, even things that feel minor.
3. Who is the audience for your work? What do they care about?
4. Voice calibration. I am going to analyze how you write, not just what you
   say about your style. Use your most important current project or
   responsibility for all three parts.
   a. Explain it to a stranger at a party with no background in your field.
      Two or three sentences.
   b. One-line status update to a colleague who already knows the context.
   c. Quick text to a friend about something that went well this week.
5. What tools do you use every day? Email, calendar, CMS, databases, code
   editors, anything. Include the ones you wish worked better.
6. What are your top 3 to 5 priorities for the next few weeks?
7. How do you work? Deadlines or no deadlines? Sprint and rest, or steady
   pace? How do you decide things when information is incomplete?
8. What should your assistant never do or say? Think about bad AI interactions
   you have had and what made them bad. This is the most important question.
9. Anything else I should know? Preferences, quirks, context that would not
   come up from the other answers?

Step 4. Generate two files from the interview.

Write assistant/CLAUDE.md with these sections: Identity, Projects and
Publications, Audience, Output Defaults, Tools, Current Priorities, Working
Style, Recognized Commands. Use the person's actual words. No placeholder text.
No generic examples. For Recognized Commands, list: /distill (read
skills/distill.md and follow the workflow), /save-state (read
skills/save-state.md and follow the workflow), /catchup (read
skills/catchup.md and follow the workflow).

Write assistant/voice_profile.md by analyzing HOW they wrote across all answers.
Do not rely on what they said about their style. Look at what they actually did.
Be specific. Vague entries are useless. Include:
- Tone axes: directness (1-10), formality (1-10), humor frequency. Add a
  one-line observation for each. Example: "Directness: 9/10 -- leads with the
  point in every answer, no throat-clearing."
- Register analysis: what changed between their party explanation, colleague
  update, and friend text. Note vocabulary shifts, sentence length, jargon use.
  Example: "Drops jargon entirely for general audiences. Sentence length cuts
  in half for the friend text. No register shift is awkward -- they do this
  naturally."
- Communication profile: vocabulary sophistication, sentence length patterns,
  whether they lead with the point or bury it, tendency to hedge or assert,
  observed education level and communication strengths or weaknesses. Example:
  "Strong precision. Technical background evident. Undersells context for new
  audiences but sharp with peers."
- Banned patterns: punctuation they avoided (example: never used em dashes --
  always a period or comma instead), hedging phrases they never reached for
  (example: "it's worth noting," "to be fair"), corporate vocabulary they stayed
  away from (example: "leverage," "stakeholders," "circle back"). These are
  things the assistant should also never produce.
- Voice anchor: one sentence that captures how they sound at their best,
  synthesized from the strongest answers. Make it a usable writing instruction,
  not a description. Example: "Write like a senior NCO filing a deficiency
  report -- direct, factual, no filler, solution at the front."
- Two or three few-shot examples drawn verbatim from their interview answers,
  labeled with which question they came from.

Write three core skill files. These provide session continuity. Without them
the system has no memory between sessions.

assistant/skills/distill.md -- End-of-session knowledge capture. Protocol:
review the conversation, identify stable patterns worth keeping (not one-time
exceptions), propose specific updates to memory/MEMORY.md with reasoning, let
the user approve or reject each one, write approved updates to the file.

assistant/skills/save-state.md -- Session state preservation. Quick mode is the
default: write assistant/state.md in parallel without running /distill, completing
in under 30 seconds. Use /save-state --full to run /distill first when a full
consolidation is needed. State file sections: Saved (today's date), What we were
working on, Decisions made, Resume instructions, Next steps. On next session start:
if state.md exists and is recent, read it, say "Resuming from where we left off,"
and follow the Resume Instructions.

assistant/skills/catchup.md -- Context reload. Protocol: read the operating brief,
the time-sensitive table, state.md if it exists, call memory_session_start to
surface relevant memories, and read todo.md. Summarize what is known and confirm
ready to continue. Use when a session has run long or after starting fresh without
state.md. The full project registry loads on demand, not during catchup.

Step 5. Show both drafts to the user. Ask what is wrong, what is missing, and
what should change. Make the edits they request. Repeat until they approve both
files.

Step 6. Tell the user they are set up. CLAUDE.md is their operating brief.
voice_profile.md is their voice calibration. The three files in skills/ are
their continuity tools -- without them each session starts cold. At session start,
read the state file and time-sensitive table, then call the memory server session
start tool to surface relevant context. Read voice_profile.md on demand when
writing in the user's name, not at every session start. Run /distill at the end
of any session where you learned something. Use /save-state when closing out.
Use /catchup if context fades during a long session.

Tell them: after 10 to 15 sessions of use, run a refinement interview. By then
you will have made real mistakes and received real corrections. That data is more
valuable than anything from a first session. A refinement interview rebuilds
voice_profile.md with evidence from actual use. They can trigger it by saying
"run a refinement interview."

Point them to https://build-tammy.netlify.app/grow.html for what to build
next.

📋 Copy

Let's build your CLAUDE.md. I'll ask you questions. Answer in plain language.
Don't overthink it. I'll generate the file when we're done.

1. What's your name and what do you do? Give me the short version and the
   longer version. The short version is what you'd say at a party. The longer
   version is what you'd put in a bio.

2. What projects, publications, or responsibilities does your work involve
   right now? List them. Include anything ongoing, even if it feels minor.

3. Who is the audience for your work? How do they talk about what you do?
   What do they care about?

4. Voice calibration. I am going to analyze how you write, not just what you
   say about your style. Use your most important current project or
   responsibility for all three parts.
   a. Explain it to a stranger at a party with no background in your field.
      Two or three sentences.
   b. One-line status update to a colleague who already knows the context.
   c. Quick text to a friend about something that went well this week.

5. What tools do you use every day? Email, calendar, documents, a CMS,
   databases, code editors, anything. Include the ones you wish worked better.

6. What are your top 3 to 5 priorities right now? Not your long-term goals.
   What actually needs to happen in the next few weeks and months?

7. How do you work? Deadlines or no deadlines? Do you sprint and rest, or
   keep a steady pace? How do you make decisions when information is incomplete?

8. What should your assistant NEVER do or say? Tone, topics, or behaviors you
   want blocked. This is the most important question. Think about a bad
   AI interaction you've had and what made it bad.

9. Is there anything else I need to know to work with you effectively?
   Preferences, quirks, context I wouldn't have from the other answers?

Take your time. When you're done, I'll draft your CLAUDE.md and voice_profile.md
and we'll review both together.