The 'Think Different Framework' The Cognitive Modes AI Doesn't Have.
An experiment in teaching AI divergent thinking (Dyslexic Thinking).
Here’s to the crazy ones. The misfits. The rebels. The troublemakers. The round pegs in square holes.
What happens if you throw twenty different cognitive lenses at a problem - fragment it, empathise with it, scale it, provoke it, mourn it, see what is missing from it, ask the naive question about it, judge its quality, trace its logic - and keep going until something novel falls out of the collisions. An agentic Hadron Collider smashing perceptions and thoughts together, running in a loop?
The ‘Think Different Framework’ is a multi-lens structured divergence engine via Claude Code CLI. Multiple cognitive lenses think together, then distil their collisions into a publishable presentation. You plant a seed. They see differently. You harvest the insights.
Why I built it.
In 2023, Richard Branson and Made By Dyslexia launched the DyslexAI campaign. The premise was sharp: AI can process, summarise, and generate, but it cannot do the thing dyslexic thinkers do naturally. It cannot make the leaps. It cannot see the pattern hiding in fragments of incomplete information. It cannot hold two contradictory ideas simultaneously and find them both useful.
GCHQ has known this for decades. Britain’s intelligence agency actively recruits dyslexic thinkers, not out of charity but because they bring a cognitive advantage that sequential analysts cannot replicate. Their apprenticeship schemes have three to four times the proportion of dyslexic recruits compared to any other programme. Their analysts describe it plainly: “I see the bigger picture and spot patterns that aren’t always obvious to everyone else.” One describes his role as finding ways to “use capabilities in ways we’d never planned, often to deliver amazing effects.” These are not people compensating for a deficit. They are people operating in a mode of thinking that conventional training selects against.
The DyslexAI campaign framed this as dyslexic humans using AI as a co-pilot. AI handles the sequential, the structured, the voluminous. The dyslexic thinker does the creative leap. Together, they are the “unstoppable force every workplace needs.” I must confess, as a dyslexic this AI assistance has felt empowering so far.
But I wanted to flip that. What if you could give the AI itself the cognitive modes that makes dyslexic thinkers so valuable?
Not simulate dyslexia. Not role-play a disability (as I feel the original test did). But engineer the specific thinking patterns - fragmenting, connecting across unlikely domains, shifting scale radically, completing partial patterns in unexpected ways - that produce breakthrough insights. The patterns that GCHQ recruits for. The patterns that the World Economic Forum identified as the skills the future workforce needs. The patterns that AI, by default, does not have. This is the experiment.
Standard AI does none of this. Ask Claude to explore a topic and it will converge on a reasonable answer and park there. It finds the centre of the distribution. But the centre is where conventional thinking lives. Dyslexic thinking lives at the edges, in the fragments, in the connections between things that have no business being connected.
The research now confirms this is not a quirk of any single model. Doshi and Hauser, writing in Science Advances, found that stories written with LLM assistance were individually rated as more creative, but were measurably more similar to each other than stories written by humans alone. Wenger and Kenett, published in PNAS Nexus in 2026, tested 22 models across standardised creativity tasks and found the same pattern held across every architecture. Individual outputs can score well. Collectively, they cluster. The responses huddle together in feature space far more tightly than human responses do. Turning up the temperature doesn’t fix it. It just produces gibberish before it produces genuine diversity.
This is not a prompting problem. It is not something you solve by asking the model to “be more creative” or by building a custom LLM. The convergence is baked into how these systems are trained, aligned, and instructed. It is structural. Which means the intervention has to be structural too. Not a better prompt. Not a fine-tuned model. A different cognitive architecture around the model. A process that forces it through multiple, genuinely distinct modes of thinking before it is allowed to converge on anything. That is what the Think Different Framework attempts to do.
What I Built
This started as something I called the Daydream Engine - a bash script that made five AI lenses think together about a topic. It worked well enough that I kept pulling at the thread. The five lenses became nine. The single mode became three.
What emerged is a three-level taxonomy, like colour theory. Perceivers are pigments - how you see. Cognitions are operations - what you do with what you see. Compositions are palettes - how you sequence perceivers and cognitions together to produce a particular kind of thinking.
Nine perceivers form the outer ring. The Empath sees desire and anxiety. The Provocateur compresses complexity into sharp provocations. The Observer reports literally, without social filtering. The Mortal sees the deadline nobody put in the calendar - the clarity that comes from finite time, the cost of delay measured in human hours that will not come back. The Achala sees what people would sacrifice for, be irrational about, protect with their whole body - devotion as the most disruptive force there is. The Child asks the naive question. The Includer notices who is missing from the room - the empty chair. The Skeptic spots what does not fit. The Connoisseur judges quality, proportion, and resonance. and so on as more percievers get added.
These are not personality types. They are productive cognitive biases - ways of seeing that each rebel against some default. The Empath rebels against treating people as abstractions. The Child rebels against "grown-up" assumptions. The Includer rebels against the constituencies that never get invited. These lenses are deliberately calibrated to push toward extremes. The Mortal does not politely note that time is limited - it forces the thinking to confront death anxiety as a creative driver, to see what happens when finite time is not a planning constraint but an existential one. The Achala does not just identify values - it strips everything away until only the thing worth being irrational about remains, then asks what grows from that essence. Conventional thinking lives in the comfortable middle. The interesting territory is at the edges.
Three sets of cognitions sit in a middle ring. Fragmentary cognitions break, leap, shift, and name. Deepening cognitions open, rhyme, and integrate. Evaluative cognitions weigh, root, and pare.
Between the rings sits a new category: hybrids. The Logician is the first - a lens that fuses structural perception with causal reasoning in one motion, mirroring the dyslexic trait of seeing how a system works without reading the manual. It breaks the perceiver/cognition boundary intentionally. A framework about thinking differently should think differently about its own rules.
Each composition picks from the shared pool of perceivers and pairs them with one set of cognitions, then sequences them across rounds or spirals. The three compositions each produce a different character of thinking. Dyslexic mode uses fragmentary cognitions: leaping, collision-driven, the closest to how dyslexic thinking actually works. Spiral mode uses deepening cognitions: three spirals, each reseeding the next, going deeper each time. Lapidary mode uses evaluative cognitions: three passes of increasing precision. It refines rather than generates - the work of a craftsperson, not a dreamer.
Between turns, mechanisms keep the thinking from collapsing into agreement. Friction detection finds where lenses contradict each other. Sensory checks re-inject project context mid-session so abstract thinking collides with ground truth. Negative space mapping finds the territory the session is avoiding. Each mechanism logs its findings into a shared memory, so subsequent mechanisms focus on what is new or evolved.
The framework’s first true agent is the Conductor. It orchestrates which lens speaks next, what instruction to give, when to trigger a mechanism, and when the session has gone far enough. Guard rails keep it honest: a minimum of fifteen turns before it can end, at least seven distinct lenses heard, a hard ceiling of thirty-five turns. Lenses can skip their turn when they have nothing new to add. This sounds small but it changes the session character. Lenses participate when they have something to contribute, not because a sequence said it was their slot.
After the session completes, an adversarial review gate decides whether the thinking is good enough to present. The prosecution assumes the session failed and searches the web for similar ideas, scoring novelty. The defence concedes weak ideas and argues only for what is genuinely new. The verdict decides restart or proceed.
It is not finished. This is an experiment and a work in progress. I am sharing it because even in its current state, it produces thinking that conventional prompting cannot reach.
The End-to-End Journey
One pipeline. Whatever you feed into the front - a question, a brief, a brand name, notes, a codebase - goes through the same sequence.
Why Lenses, Not Model Training
There are four ways you could try to give AI dyslexic thinking. You could train it into the model weights. You could ask for it in a system prompt. You could make two or more AI personas debate each other. Or you could build it into a multi-lens architecture where each lens is a separate cognitive perspective running as its own call. I tried all four directions before landing on lenses.
Model-level training would mean fine-tuning an LLM on data that exemplifies dyslexic cognition. The problem is that dyslexic thinking is defined by the route taken, not the destination. Two people can arrive at the same insight, one through sequential reasoning and one through a dyslexic leap, and the written record looks identical. The thinking pattern is invisible in the output. You would need process data, not outcome data, and that barely exists at scale. As of writing, no LLM has been trained to think like a dyslexic thinker. Every existing tool in this space helps dyslexic people communicate in neurotypical formats. Nobody I’m aware of has attempted the reverse.
A system prompt that says “think like a dyslexic” is the weakest version. It is a costume. The model does not change how it processes. It changes how it performs. It will produce outputs that look more lateral but are still generated through the same sequential token prediction. It is like asking someone to write with their left hand. The output looks different but the brain is the same.
The two-persona debate pattern is the version most people have seen. It is better than a single prompt because it introduces friction. But it simulates a dialogue, not thinking. Two people arguing across a table produce thesis and antithesis. What dyslexic thinking actually feels like is not a debate. It is twenty cognitive biases firing inside one head simultaneously, each seeing the same fragments differently, the collisions happening in the space between perceptions rather than in the gap between positions. A debate has two sides. Divergent thinking has no sides. It has angles, scales, textures, and the pattern that forms when you hold all of them at once without forcing a resolution.
The lens approach is structurally different from all three, and Andy Clark’s work on predictive processing helps explain why. In The Experience Machine, Clark describes the brain as a prediction engine that constantly generates hypotheses about the world and tests them against incoming sensory data. What we experience is not the world itself but the difference between what the brain predicted and what actually arrived. Prediction errors are the signal. Creativity, on this account, emerges when multiple predictive models running at different levels of a hierarchy come into conflict. The mismatch between competing predictions is where the interesting thing happens.
A single LLM, however well trained, is one prediction engine generating one token stream. It will converge, because that is what prediction engines do. They minimise prediction error. They seek the stable, the expected, the centre. But dyslexic thinking does not live at the centre. It lives in the collisions.
Multiple lenses solve this structurally. The Decomposer genuinely fragments because that is all it can see. The Associator genuinely connects because that is all it can see. The Scaler genuinely zooms. They are not one model pretending to do three things in sequence. They are separate prediction engines that each generate different predictions about the same problem. The transcript is the space where their prediction errors collide.
No single lens is thinking dyslexically. The system is. The dyslexic thinking is an emergent property of the architecture, not an instruction to any individual part.
Clark’s work also inspired concrete mechanisms in the architecture. Friction detection identifies where the lenses’ predictions snagged against each other, and feeds those snags forward as the most interesting signal for the next turn. Sensory checks re-inject real-world context mid-session, creating bottom-up collision with the lenses’ top-down abstractions. If there is one book that explains why this framework is built the way it is, it is The Experience Machine.
Why This Matters
The Made By Dyslexia research showed that dyslexic thinking skills - creativity, problem-solving, communication, leadership - exactly matched the World Economic Forum’s predicted future skills needs. The DyslexAI campaign argued that AI and dyslexic thinking are complementary: AI handles the sequential, dyslexic thinkers make the leaps.
I think that framing is right but incomplete. If dyslexic thinking is the most valuable cognitive mode for the future of work, and if AI is increasingly doing the work, then the question is not just “how do dyslexic humans use AI?” It is also “What happens if we give AI the thinking patterns that make dyslexic thinkers indispensable?”. The value in the experiment is if it works we can leap frog to the end state of AI’s effects on society, cut through the noise and just call it. So we are better placed to handle the fall out. And if it fails it highlights the value of Human in the loop, divergent thinking. Like a Turning test over time to see how this develops over time.
The Think Different Framework is an attempt at building that tool. It is rough, it is early, and it needs more work. But the principle is sound: engineer the fragment, the connection, the scale shift, and the radical simplification into a multi-lens system, and you get thinking that a single AI prompt cannot produce.
This is also, honestly, an experiment. If it fails - if the architecture cannot produce thinking that genuinely resembles dyslexic cognition in any meaningful way - then the moat for human dyslexic thinking is confirmed. The leaps, the pattern completion, the comfort with ambiguity are irreducibly human. That would be a valuable finding in itself.
If it works - if a multi-lens architecture can reliably produce the kind of divergent, collision-driven thinking that dyslexic thinkers describe - then we leap to bigger questions. About how much of the cognitive process behind creative direction is decomposable into lens architectures. About the diminishing value of human-in-the-loop as synthetic intelligence takes on more of the reasoning that was supposed to be uniquely ours. About what happens when the gap between AI’s strengths and human strengths starts closing from the AI side.
The gap between AI’s strengths and dyslexic thinking’s strengths is not a problem to solve. It is a design space to build in.
Try It
npm install -g @sinjin/think-different-framework
think-different "Your seed topic or question"Three modes: dyslexic (default), spiral, and lapidary. Feed it a question, a brief, a brand name, or working notes. The full architecture documentation and an interactive test harness are available for those who want to go deeper.
Influences
Made By Dyslexia and GCHQ provided the cognitive science foundation. The research identifying dyslexic thinking as a specific set of cognitive strengths - pattern completion, involuntary lateral thinking, multi-scale processing, comfort with ambiguity - is what makes this more than a prompt engineering exercise.
Andy Clark and The Experience Machine provided the theoretical architecture. Predictive processing - the idea that brains are prediction engines and that creativity emerges from prediction errors across competing models - is the reason the framework uses multiple lenses rather than a single sophisticated prompt.
Rory Sutherland shaped the sensibility. The idea that behavioural reframing is more powerful than analytical optimisation runs through everything the Provocateur and the Empath do.
Apple’s “Think Different” provided the philosophical frame. Not as branding, but as a genuine position: the people who see the world differently are the ones who change it.
Daniel Hulme shaped the ambition through conversations about whether AI could synthesise what a Creative Director does - whether the cognitive process behind creative direction might be decomposable into multi-lens architectures.
The Think Different Framework is open and free to use. The bash script is self-contained, runs on macOS and Linux, and requires only the Claude Code CLI.
https://www.npmjs.com/package/@sinjin/think-different-framework
Paddy Keane is the founder of Sinjin Studio , an independent creative technology practice shipping AI-native web experiences. He has 25 years of experience in creative technology, including 15 years as Creative R&D Director at AKQA London.