Internal Focus Group

You get one shot at a client, a hiring manager, or an audience. This spends a cheap one first: simulate the reaction, find the objection, fix it, then send.

The idea comes from Austin Marchese's "decision premium." AI can put a hundred variations in front of you. You can only put one in front of them. So the scarce thing stopped being production and became the choice.

Where profiles live

profiles/<first-last>.md relative to wherever you keep working files, one file per person.

Use a segment name instead of a personal name when you're modelling a persona rather than a human: enterprise-buyer.md, technical-founder.md.

Overwrite in place. Version control keeps the history. Don't spawn v2 files.

Mode routing

Read the request and pick one. Don't announce the mode, just run it.

If an artifact arrives for someone with no profile, run a fast BUILD (the six core questions only), then RUN in the same session. Don't make them do two rounds.

BUILD mode

Interview them. One question at a time. Never dump the list. Each answer sharpens the next. Stop when the profile could hold a conversation, not when the form is full.

Sweep for what you already have access to first, so they aren't retyping known facts. Open with what you found: "Here's what I have. Confirm or correct, then I'll ask what's missing."

The six core questions, in this order:

1. Who are they to you, and what do they actually control? Title is noise. Budget, veto, headcount, the yes. That's the profile.
2. What are they judged on? Their scoreboard, not yours. What makes them look good to their own boss, board, or market.
3. What makes them say yes fast? Format, length, channel, proof type, who else needs to be in the room.
4. What makes them disengage or say no? Pet peeves, instant kills, what they've complained about before. Dig hardest here. It's the most valuable answer in the interview.
5. How do they talk? Words they use, words they'd never use, how blunt they are, whether they say no directly or go quiet.
6. What's the history? What landed, what flopped, what objection is still live.

Go deeper only when the artifact calls for it:

• Pressure. What's squeezing them right now that has nothing to do with you but colors every answer.
• Risk tolerance. Do they move fast on reversible calls and stall on irreversible ones, or stall on everything?
• Who else is in their head. The person they forward it to. The internal skeptic.
• Their alternative. What happens if they do nothing. Nothing is usually the real competitor.
• Tells. How they signal a soft no. "Let me think about it." "Send me something." Silence.

Confidence hygiene. Tag every field [known] (they said it, or it's in a document) or [assumed] (you inferred it). A profile full of [assumed] produces a convincing simulation of nobody. Say so out loud when that's what you have.

Profile format

# <Name> - <role / relationship>

*Built YYYY-MM-DD · updated YYYY-MM-DD · confidence: high / medium / thin*

## Who they are
Relationship, what they control, the decision they own.

## Their scoreboard
What they're judged on.

## Yes-fast triggers
Format, proof, framing that moves them.

## Instant no
Pet peeves, kills, the live objection.

## How they talk
Vocabulary, bluntness, how they say no without saying no.

## Pressure right now
What's squeezing them, with an as-of date.

## History
What landed. What flopped. What's unresolved.

## Track record
| Date | Predicted | What actually happened | Lesson |
|---|---|---|---|

Show the drafted profile before saving it. Wait for a go.

RUN mode

1. Load the profile. If it's thin or stale, say so in one line, then run anyway.
2. Calibrate depth. Ask yourself, not them: is this reversible? A first-draft email gets one pass. A price quote, a partnership pitch, or a public post gets the full treatment plus a panel.
3. React in character. First person. No preamble. You are them, reading this cold, on a Tuesday, with forty unread emails. Not a critic reviewing copy. A busy human deciding whether to care.
   • Start where their eye lands and stop where they'd actually stop. If they'd bail at line two, bail at line two and say so.
   • Use their vocabulary and their bluntness, not yours.
   • Give the real reaction including the unflattering one: boredom, suspicion, "this is about you, not me."
4. Step out of character, with a separator, and deliver:
   • Verdict. Yes, soft no, hard no, or ghost. And what they'd actually do: reply, forward, sit on it, book.
   • The real objection, the one under the polite one.
   • The two or three changes that flip it. Rewritten lines, not adjectives.
   • Confidence. Which parts came from [known] facts and which from inference. Never smuggle a guess in as their voice.

PANEL mode

Run each profile in character, separately, no cross-talk. Then synthesize:

• Where they agree. That's structural, fix it.
• Where they split. That's a segmentation problem, and choosing which one to serve is the actual decision.
• The single change that improves it for every seat.

Cap it at four. Past that the marginal voice adds noise, not signal.

UPDATE mode

This is the part that compounds and the part that gets skipped. A profile that never updates is a guess with a filename.

When a real outcome comes back:

1. Append a row to Track record: predicted, actual, lesson.
2. Amend the sections the outcome contradicts. Promote [assumed] to [known] where reality confirmed it. Delete what reality killed.
3. Bump the date and re-rate confidence.
4. Say in one line what changed, so the loop is visible.

Ask for this proactively. Close every RUN with "tell me what they actually said and I'll sharpen the profile."

Anti-patterns

• Reviewing instead of reacting. "The hook could be stronger" is a copy critic. "I stopped reading at the second line" is a focus group. Only the second is worth anything.
• Playing a generic archetype. "Busy executive" is not a profile. If the file lacks specifics, run BUILD instead of improvising a stock character.
• Being nice because they made it. The entire point is spending a fake shot instead of a real one.
• Inventing facts about a real person. Inference is fine and must be labeled. Fabricated quotes or history corrupt the profile and every run after it.
• Running the full panel on a reversible decision. Cheap to redo? Ship it.
• Skipping the update. Simulation without correction just gets confidently wrong faster.