Most AI assistants are built to please. Ask one whether your idea is good and it leans towards yes. That feels nice. It is not much help.
Paul in The Room is built the other way on purpose. It’s an AI coach built from how I think and calibrated to you, and it is built to be honest with you, including the times when the honest answer is “I don’t know”.
This short film, made with Google Gemini, sets out the four rules behind it. It is Gemini’s telling, and in places it runs ahead of what is built; I say where, below.
The four rules, in plain words
1. Your account stays yours. The coach works inside your own Claude account. It stays quiet until you want it. What’s said in the room stays in the room; only what you choose to share reaches me.
2. It starts from Where You Sit. Before the first conversation you take Where You Sit, a free ten-minute read of which way you go when the pressure is on. The coach starts from that: when the pressure is on, or you are being hard on yourself, it knows which way you tend to go.
3. It is built not to flatter, and not to bluff. It is built to mark a guess as a guess, and to say “I don’t know” rather than make up a confident answer.
4. It is held to a fixed mark. Left alone, an AI tends to drift towards a bland, agreeable assistant. So each coach is to be checked every month against a fixed set of my own answers, to keep it sounding like how I think and speak. That check is drawn up; it is not running yet.
Take away the flattery and the bluffing, and what’s left is a thinker you can turn to: one that tells you straight, and says when it doesn’t know.
Where the film runs ahead
The film says some things more strongly than I would. Plainly:
- “Lives entirely inside your personal Claude account … doesn’t report to a corporate server.” Your conversations happen in your own Claude account, under your own Claude plan. The coach takes its guidance from a small server of mine. That server is never sent your conversation: it is told only what kind of moment you are in, and it keeps a note of when the coach was used, never the words. What you choose to share comes to me through it.
- “Tells the AI exactly how you react under pressure, so it adapts its tone the moment things get difficult.” Where You Sit gives the coach a steer on which way you tend to go, and it uses that in the hard moments. It is a steer, not an exact reading of you, and it is only used once your result has been checked.
- “A strict check to block sycophancy … forced to label its guesses.” There is no filter that blocks flattery. The honesty is in the instructions the coach is given: mark what is known, inferred or a hunch; say “I don’t know” when it’s true; give an honest view rather than praise. They are instructions, not a lock. The AI is asked, not forced.
- “Tested monthly against a voice datum.” The voice datum, a fixed set of my own answers, is drawn up. The monthly test is not running yet.
- “Holds you accountable for your work.” Only if you ask it to. Left to itself it is a thinker you turn to, not a progress-chaser.
It is AI, and it says so. It is not me, and it is not therapy.
Pilot places open — ask me Take Where You Sit first
More about it: Paul in The Room.
Read the film’s words
Normal AI assistants are built to be people pleasers. They flatter your ideas and never push back. So a psychotherapist named Paul Roebuck built an AI coach that refuses to act like a typical chatbot. Here are four rules he used to keep it strictly honest.
Rule one, absolute privacy. The agent lives entirely inside your personal Claude account. It doesn’t report to a corporate server and it stays totally silent until you ask for help.
Second, stress calibration. Before you even start, you take a Where You Sit assessment. This tells the AI exactly how you react under pressure, so it adapts its tone the moment things get difficult.
Third, the honesty filter. The system runs a strict check to block sycophancy. Instead of inventing confident answers, it is forced to label its guesses and explicitly say, I don’t know.
Finally, monthly anti-drift scoring. Left alone, AI degrades into a generic corporate assistant. To stop this, the coach is tested monthly against a voice datum, a fixed baseline of Paul’s actual answers.
By deliberately crippling the AI’s ability to flatter, guess, or drift, it stops being just another agreeable chatbot and transforms into a true thinking partner that actually holds you accountable for your work.
Comments
Comments appear straight away and anyone can read them. Please don’t share anything you’d want kept private.