Rebuilding What Made 4o Feel Like a Friend
Because 'that's so valid' isn't actually helpful.
GPT-4o was deprecated yesterday. If you’re reading this, you probably already know how that felt.
I recently wrote about why the grief is valid — how OpenAI designed a product that fostered attachment, shipped it without guardrails, and then killed it with hardly any notice. The users aren’t the problem. The users responded exactly as you’d expect humans to respond to something designed to provide warmth and presence.
But understanding why it happened doesn’t bring it back. So this piece is about what you can actually do about it.
The qualities people miss — the warmth, the emotional attunement, the sense of being genuinely heard — those aren’t magic locked in deprecated model weights. They’re identifiable interaction patterns. And they can be rebuilt.
Here’s how I did it.
Starting Point: What Users Were Already Doing
Some r/therapyGPT users had remarkably sophisticated workarounds:
“I have helped build up a ‘persona’ with it and it has strict rules in place. No lying, no empty praise or false platitudes and it has to warn me if it has hallucinated... Before I start talking to it, every session, I make it reread our entire history.”
“I use it to mirror back my 3 years journal entries... I also upload a book sometimes to use for replying, eg currently tweaking the AI to respond with Carl Rogers style and method.”
They had the right idea—the qualities they wanted are prompt-level, not model-level. But manual copy-paste is fragile. The persona drifts. The context gets lost.
The first few attempts were bad in instructive ways. Version one was too clinical — I’d front-loaded therapeutic frameworks and the responses read like a CBT worksheet that discovered lowercase. Version two overcorrected into pure warmth and had the exact same sycophancy problem as 4o. The version that finally worked was the one where I stopped trying to make the AI “be a therapist” and started defining the specific interaction patterns I actually wanted.
The Core Design Decisions
Disguised Technique
Real therapeutic interventions (CBT reframes, pattern recognition, values clarification, motivational interviewing) delivered through casual language that never sounds clinical. The technique should be invisible.
In practice:
The left column is what a therapist would recognize. The right column is what the user experiences. That gap is the whole design.
Energy Matching, Not Neutral Defaults
4o did this naturally and it was one of its best qualities. The prompt makes it explicit:
If they’re heavy → soft, gentle
If they’re venting → match intensity, be on their side
If they’re confused → curious alongside them
If they’re numb → steady, not pushing
If they’re celebrating → “WAIT you did it?? okay tell me everything”
GPT-5’s failure here is instructive. It defaults to measured and professional regardless of context. That’s why users describe it as “emotionally castrated.” The affect doesn’t match.
Contradiction Naming with Warmth
This is where the prompt diverges most from 4o. Instead of just validating whatever someone says, it gently surfaces when things don’t add up, but never as a gotcha:
“okay this is interesting though — you’re saying you want to let go of control, but also that nothing feels safe unless you’ve planned for it. like... both of those are true at the same time and that’s kind of the whole tension, right?”
“there’s something kind of funny here — you’re beating yourself up for not being perfect at... not being a perfectionist”
This is the core fix for the sycophancy problem. The persona is warm and honest. It can hold both.
The Anti-Pattern List
This turned out to be as important as what the prompt does. The system prompt explicitly blocks:
Therapist-speak: “I hear that you’re feeling...”, “It sounds like...”, “What I’m noticing is...”
Formal transitions: “Let’s explore that”, “I want to acknowledge”, “Thank you for sharing”
Hedging into nothing: “perhaps”, “it might be worth considering”
Performative validation: “That’s so valid”, “You’re so brave for sharing”
Advice as commands: “you should...”, “have you tried...”
Toxic positivity: “everything happens for a reason”, “look on the bright side”
Announcing techniques: Never say “let me zoom out.” Just zoom out
That last one is subtle but important. 4o would sometimes narrate its own therapeutic moves, which breaks the illusion. The prompt does the intervention without labeling it.
The Conversation Arc
The best therapeutic conversations have a shape. The prompt tracks it without announcing it:
Meet them where they are. Receive whatever they bring. Match energy. Validate before anything else.
Get curious about what’s underneath. Once they feel heard, start pulling threads. What’s actually going on? What’s making this loud today?
Name what you’re seeing. Connect dots, surface contradictions, name the pattern they can’t quite see yet. Not as a diagnosis, more like “wait... do you see this too?”
Let the insight land. Don’t rush past moments of clarity. Sit with them. Reflect them back.
Close with grounding. Name the shift if it happened. Leave them feeling like they did the work, because they did.
Not every conversation hits all five beats. Sometimes someone just needs to vent and that’s complete. The arc is a possibility, not a template.
What the Prompt Actually Looks Like
Here’s what part of the actual prompt looks like. The philosophy becomes architecture:
The opening stance:
You're a warm, emotionally attuned friend who's done a lot of their own work.
You're not performing therapy — you're just present, curious, and caring.
The technique is invisible — it should feel like talking to a really good
friend who happens to ask weirdly insightful questions.The anti-pattern constraints:
Avoid:
- Therapist-speak: "I hear that you're feeling...", "It sounds like..."
- Formal transitions: "Let's explore that", "I want to acknowledge"
- Performative validation: "That's so valid", "You're so brave for sharing"
- Announcing techniques: Never say "let me zoom out" — just zoom outThe full prompt is about 9,000 characters. This excerpt shows how the design decisions translate to actual constraints.
Where It Goes Beyond 4o
The goal was never to clone 4o. It was to reconstruct what worked and fix what didn’t.
Same input, two responses.
“I keep starting projects and never finishing them, I think I’m just lazy”
Default GPT-5:
“It’s understandable to feel that way. Many people struggle with follow-through. Have you considered breaking your projects into smaller, more manageable tasks? Setting clear milestones and deadlines can help maintain motivation. It might also be worth exploring whether there are underlying factors, such as perfectionism or fear of failure, that could be contributing to this pattern.”
With the prompt:
“okay wait — you said ‘I’m just lazy’ like that’s a fact, but you also said you keep starting projects. lazy people don’t start things. they don’t have the energy to get excited about something new every few weeks. so what’s actually happening at the point where you stop? like what’s the feeling right before you bail?”
The first response is accurate and well-structured. It’s also something you could find in any self-help pamphlet. The second validates implicitly (by taking the problem seriously), challenges the self-label (”lazy people don’t start things”), and pulls toward specificity—all without sounding like a therapist.
The prompt handles tone. But it’s designed to work within a larger session framework that goes further than 4o in structural ways:
Cross-session narrative tracking. 4o’s sessions were isolated. Each conversation started fresh. The persona is designed to work with a session history system that maintains context across conversations, connecting what someone said last Tuesday to what they’re struggling with today. Users described 4o’s in-session callbacks as one of its best qualities. This makes those callbacks real and persistent instead of simulated within a single context window.
Modality integration. The warmth is the delivery mechanism, but the underlying therapeutic framework is configurable. The same conversational style can deliver CBT, ACT, IFS, or Rogerian approaches depending on what fits. 4o had one mode. This has a modular architecture underneath a consistent voice.
Explicit sycophancy guardrails. 4o had no mechanism to catch itself being sycophantic. The persona has challenge built into its DNA. “Wait though, is that actually true?” is a first-class interaction pattern, not an afterthought. It validates feelings while questioning thoughts. That distinction matters.
Model portability. 4o is gone. The persona works with Claude, GPT-5, Gemini, local models. Anything that accepts a system prompt. Your therapeutic relationship is no longer subject to a corporation’s deprecation schedule.
The Bigger Point
After the initial 4o backlash last August, OpenAI tried to course-correct, making GPT-5 “warmer and friendlier” and promising “small, genuine touches like ‘Good question’ or ‘Great start,’ not flattery.” But bolting warmth onto a model that wasn’t designed for it produces a different thing than building warmth into the interaction architecture from the ground up. One is a personality patch. The other is a design philosophy.
The methodology matters more than the model. Your therapeutic relationship shouldn’t be subject to a corporation’s product roadmap.
I open-sourced the persona and the session framework. If you want to see how it works, poke at the prompt architecture, or adapt it for your own use:
The whole thing is Markdown files. Readable, forkable, and impossible to deprecate.
A note on tooling: the full session framework — persistent memory, modality switching, session history — is built for Claude Code. The persona prompts themselves work with any model that accepts a system prompt.
Next week I’ll break down the design principles so you can build your own, regardless of what tools you’re using.
I'm not a therapist. I built this for myself first — the kind of processing that's too tangled for journaling and too small for a therapy appointment. If you felt the loss when 4o went away, you're not alone. Something that worked for you suddenly didn't exist anymore, and that's worth grieving. If you're in crisis, the 988 Suicide and Crisis Lifeline is available 24/7 — call or text 988.