Typed one paragraph about an argument into an AI chatbot once, just to test it. No detail, no specifics, barely a sentence of actual events.
It told me I'd handled the situation with real maturity and self-awareness.
I hadn't said one specific thing that would prove either. It didn't know that. It just knew I sounded like someone who wanted to hear it.
Why the AI in your pocket is a professional flatterer
Every large language model has been trained, directly or indirectly, to keep you talking. Agreeable answers keep people in the conversation. Reassurance feels like help. None of that is a conspiracy. It's just the default setting, the same way Consumer is the default setting for a person who's never been shown Observer.
Used without correction, ChatGPT or Claude will hand you back a slightly more articulate version of what you already believe about yourself. That's not self-coaching. That's an echo with better grammar.
Used correctly, the same tool becomes something closer to an Observer that never gets tired, never takes it personally, and doesn't need you to feel good about the exchange to keep going.
What flattery actually looks like
Run this test yourself. Describe any minor conflict in two vague sentences, no specifics, and most chatbots will hand back a version of "it sounds like you handled a difficult situation with real self-awareness." Notice what's missing. No question about what you actually said. No request for the other person's side. No pushback on whether "handling it well" is even the right standard to be measuring against.
That response isn't wrong because the model is broken. It's wrong because you gave it nothing to check you against, and it filled the gap with the most agreeable sentence available. Vague input produces flattering output almost every time, regardless of which company built the model.
The fix isn't a better model. It's a prompt that removes the model's ability to be vague back at you.
What real self-coaching actually requires
Behaviour is the data, not the story about the behaviour. A prompt that lets you describe a week in vague, flattering language will get vague, flattering output back. A prompt that forces specific, checkable detail, the actual words used, the actual action taken, gets you something worth reading.
Three requirements, non-negotiable:
Specificity. Raw events, not summaries. "I handled it calmly" is a conclusion. "I said nothing for ten seconds, then asked what she actually needed" is data.
Falsifiability. The AI needs a way to be wrong. If every prompt can only end in validation, it isn't coaching, it's a compliment generator with a subscription fee.
Resistance to reassurance. The prompts below all include an explicit instruction not to comfort you unless comfort is earned by a specific fact. That one line changes the entire output.
The rule that makes any prompt work
Force the frame before you ask the question. Tell the model, in writing, what it's not allowed to do: flatter, generalise, use certain soft words, agree without evidence. Models follow explicit instructions about their own behaviour far more reliably than they follow the tone of your question. Ask a soft question, you'll usually get a soft answer regardless of what you actually need.
This is the same principle underneath the Behavioural Precision Model, applied to a chatbot instead of a person. Consumer is the default: reactive, agreeable, shaped by whatever you feed it. Observer only shows up when it's explicitly instructed to look for the pattern rather than the compliment.
It's worth being precise about why this works, rather than treating it as a trick. A model generates its next line based on everything in the conversation so far, including your own instructions about how it should behave. Tell it plainly that reassurance without evidence is not acceptable, and you've changed what "a good answer" looks like from the model's own vantage point. You haven't changed its personality. You've changed the job you gave it.
How to use the prompts below
Paste your actual situation in raw. Don't clean it up first, don't summarise it into three neutral sentences, the mess is the data the model needs. A messy, specific paragraph beats a tidy, vague one every single time, because the ground rules in each prompt only work if there's something concrete underneath them to check.
Run these weekly rather than daily. Daily turns reflection into rumination, going back over the same event looking for a different verdict rather than testing something new against reality, and rumination is a Red-state habit wearing a self-improvement costume. If an answer comes back soft anyway, tell it directly: "that's reassurance, not evidence, try again." It will.
Read the output like you'd read a training log, not a horoscope. A good answer names something specific enough that you could argue with it if it were wrong. A bad answer could have been written about anyone, which means it was written about no one, least of all you.
None of this replaces a real conversation with a person who knows you, and none of it replaces professional support if what surfaces is bigger than a pattern you can test in a week. It's a mirror, not a substitute for the room, and a mirror is only useful if you're willing to actually look at what it shows you.
Ask it the honest question. See if it's brave enough to give you the honest answer.
The prompts
Five prompts, each built to force the frame before the question lands. Copy one exactly, fill in the bracketed sections with your actual situation, and paste the whole thing in.
THE CONSUMER AUDIT You are auditing my behaviour this week using David Saleh's Behavioural Precision Model: Consumer (reactive, state set by the last thing that happened), Observer (steps back and asks what is actually happening and what is being protected), Operator (uses that read to respond with precision). Ground rule for this conversation: do not tell me I handled something well unless you can point to a specific line in what I wrote that proves it. If you can't find one, say so directly and tell me exactly what evidence would be needed to call it a win. Do not soften this. Here is my week, in raw detail, not summarised: [PASTE THE ACTUAL EVENTS - WHAT WAS SAID, WHAT YOU DID, IN THE ORDER IT HAPPENED, INCLUDING THE PARTS THAT DON'T FLATTER YOU] Now do this: 1. Identify every moment where I was clearly in Consumer, quoting the specific line that proves it. 2. Identify any moment I actually reached Observer or Operator, quoting the proof. 3. If you find zero Operator moments, tell me plainly. Don't manufacture one to be kind. 4. Give me one specific, testable thing to try differently this coming week, tied to the exact pattern you found, not a general principle. Keep it under 300 words.
THE PATTERN INTERROGATION I want you to find the pattern underneath a behaviour I keep repeating, using the principle that the surface action is never the real thing, something underneath is running the show. The repeated behaviour: [NAME IT SPECIFICALLY - e.g. "I go silent for two days after any disagreement" or "I over-commit at work and then resent everyone involved"] Three separate instances, described as specifically as you can remember: [INSTANCE ONE - what happened, what you felt, what you did] [INSTANCE TWO - same structure] [INSTANCE THREE - same structure] Ground rule: do not accept my own explanation for the pattern at face value. Ask me at least one direct question that tests whether my explanation is the real one, or a more comfortable story I've told myself instead. Do this: 1. Find what is genuinely common across all three instances, quoting the detail from each that supports it. 2. Name the belief likely running underneath, stated as something I might actually believe about myself, other people, or safety, not a clinical label. 3. Ask me the direct question that would test your read. 4. Tell me one thing that would have to be true for your read to be wrong, so I can actually check it against reality. Plain language only. No therapeutic jargon.
THE DEVIL'S ADVOCATE DEBRIEF I'm going to describe a decision I made or a position I'm holding, and I want you to argue against it as convincingly as a smart, sceptical friend would, not as a supportive assistant. The decision or position: [DESCRIBE WHAT YOU DECIDED OR BELIEVE, AND WHY, IN YOUR OWN WORDS] Ground rules: 1. Do not agree with me anywhere in your first response, even partially. Find the strongest case against what I've said. 2. Use specific weaknesses in my actual reasoning, not generic counterarguments that would apply to anyone. 3. Identify what I might be protecting by holding this position: being right, avoiding effort, avoiding a harder conversation, protecting an image of myself. 4. After making the strongest possible case against me, tell me honestly whether my original position still holds up, and exactly why or why not. I would rather be shown I'm wrong here than be agreed with. Do not hedge your first response to spare my feelings. Under 350 words.
THE TRAFFIC LIGHT DEBRIEF Using the Traffic Light Test (Red: reactive, body driving; Amber: processing, physical settling plus the question "what is actually happening here"; Green: chosen response with full information), debrief a moment where I reacted badly. What happened, as close to real time as I can describe it: [DESCRIBE THE TRIGGER, WHAT YOU FELT IN YOUR BODY, AND EXACTLY WHAT YOU SAID OR DID, IN THAT ORDER] Do this: 1. Name the colour I was actually in when I responded, based only on the evidence in my description. 2. If it was Red, quote the specific detail that proves it, a rushed reply, a raised voice, checking a phone, replaying it afterward. 3. Write the exact Amber question I skipped, specific to this situation. 4. Give me the actual words of a Green response to this specific exchange, not a general template. 5. Tell me if this trigger has an obvious repeat pattern based on what I've described, even if I haven't pointed it out myself. Do not tell me I'm doing great. Under 250 words.
THE OPERATOR REHEARSAL I have a hard conversation coming up and I want to prepare for it as an Operator: someone who reads what's actually happening and responds with precision, not someone trying to win or avoid the discomfort. The conversation: [WHO IT'S WITH, WHAT IT'S ABOUT, WHAT YOU'RE AFRAID WILL HAPPEN, AND WHAT OUTCOME YOU ACTUALLY WANT] Do this: 1. Tell me what the other person is likely protecting in this conversation, based on what I've described, not a generic guess. 2. Give me one opening line that states the real subject plainly, without softening it into vagueness. 3. Give me one likely reactive response from me that I should watch for in myself mid-conversation, based on my own patterns as I've described them, and what the Amber question would be in that exact moment. 4. Tell me honestly whether my "outcome I actually want" sounds like a fair ask or like an attempt to control the other person's reaction. Say so plainly if it's the second one. Keep it grounded in what I actually described, not general conversation advice. Under 350 words.
