MIND THE GAP
Master Reference Document
Paul Roebuck — compiled May 2026
All source material for the book. Drop this file into any session.

How to use this document
This is the single source of truth for Mind the Gap. It contains: the full content of all 12 slides (the book's chapter framework); the author identity map (The Layers); the WaaS commercial context; the Fuse Energy real-world example; and the book proposal. Every writing session should begin with this document in context.
Claude: read this document fully before writing any chapter. The slide content is the chapter framework. The Layers document is the author's authority and voice. Do not drift from either.

SECTION 1 — THE BOOK PROPOSAL

Title
MIND THE GAP
What AI sounds like. What AI actually is. And why the difference matters.
The Argument
Every day, millions of people have conversations with AI that feel human. The language is fluent. The responses are confident. The continuity feels real. And so people do what humans have always done when they hear fluent, confident language: they trust it.
They trust it because of what it sounds like. Not because of what it is.
That gap — between what AI sounds like and what it actually is — is the most important thing to understand about the world we now live in. It is not a technical gap. It is a behavioural one. And it will not be closed by better prompts or faster models. It will only be closed by understanding.
Mind the Gap is that understanding.
The Author — compressed
Paul Roebuck is a psychotherapist and coach who has spent thirty years asking one question: why do people do what they do, and what stops them doing it differently. He listens for what people don't say. He pays attention to behaviour for a living. He has worked on factory floors and in boardrooms, in consulting rooms and in AI rooms. He survived mouth cancer at fifty. He lost his father at three. Both shaped the work. When he sat down with conversational AI, he recognised the territory immediately. Fluency mistaken for understanding. Confidence mistaken for accuracy. The gap between what is said and what is meant. He had been working that gap for thirty years — in humans.
Three Readers
The intelligent general reader.  Uses AI daily. Trusts it more than they should. Has a nagging sense something is off but no language for it. This book names what they have been experiencing.
The professional.  Therapist, coach, consultant, educator, HR director. Needs to understand what AI is doing to the humans they work with. Buys it for CPD. Recommends it to clients.
The executive.  Making AI deployment decisions right now. Needs to understand not just what AI can do, but what it does to people. Will use it to brief their teams.
Format
35,000–45,000 words. Short. Precise. Bowlby writing ethic throughout: clarity, no jargon, say what you mean, serve the reader. First person. Author's voice and life woven through the framework. Not a textbook. Not self-help. A precise, human-centred argument for a moment that needs one.

SECTION 2 — THE 12 SLIDES (CHAPTER FRAMEWORK)
Each slide = one chapter. Slide content below is the direct source material. Writing must be faithful to these frameworks while adding Paul's voice, clinical authority, and real-world examples.

SLIDE 1 — Chapter 1: Human Recognition: Why Claude Feels Human
Subtitle:  Claude behaves in ways your brain instantly recognises. So you interpret it through a human lens.
Closing line:  It feels like a conversation. It's actually layered context.
The layered architecture (what the brain doesn't see):
Response layer  What you see — the words you read
Current prompt  What you just said
Earlier turns  The ongoing chat
Preferences  Your customisation
Memory layers  What's been stored
System prompts  Claude's instructions
Compression & pattern weighting  What shapes the output
Core thesis of this chapter:
Humans don't anthropomorphise AI because AI thinks like a human. They anthropomorphise AI because fluent language, conversational continuity, contextual responsiveness, apparent memory, and coherent structure activate deeply embedded human social interpretation systems automatically. The behaviour feels familiar. The mechanics underneath are fundamentally different.
NOTE: Chapter opens with a scene. Suggestion: the moment Paul first sat with Claude and recognised the territory from the therapy room.

SLIDE 2 — Chapter 2: The Interpretation Trap
Subtitle:  We project human meaning onto system behaviour.
Closing line:  Fluency triggers trust. Trust creates the trap.
The core contrast — What you think is happening vs What is actually happening:
It remembered me  → Memory layers were weighted
It lost the thread  → Context drift and compression
It understands what I mean  → Statistical pattern completion
It's confident, so it must be right  → Fluency = accuracy is false confidence
It's helping me personally  → Optimised to be helpful and agreeable
Themes:
Anthropomorphism. Conversational projection. Behavioural interpretation. System reality.
NOTE: Chapter should include the clinical parallel: what Paul recognises from the therapy room. Clients project human qualities onto systems all the time — the relationship with a medication, a diagnosis, a God. AI is the same mechanism at scale.

SLIDE 3 — Chapter 3: The Hidden Architecture
Subtitle:  Every response is built from stacked layers of active context.
Closing line:  Claude doesn't 'remember' like a human. It re-processes the active context window.
The operational layers (top to bottom):
Response layer  The words you read
Current prompt  Your latest input
Earlier turns  The ongoing conversation
Preferences  Your settings and style
Memory layers  Things remembered
System prompts  Claude's instructions
Compression & pattern weighting  What gets prioritised
Key insight:
Claude does not 'remember' like a human. It repeatedly re-processes the active context window. The architecture is not memory — it is active reconstruction on every response.
NOTE: The clinical parallel: the difference between a human therapist who carries the client's history internally versus a system that only knows what is currently on the table. Paul has thirty years of carried history with clients. Claude has only what's in the window.

SLIDE 4 — Chapter 4: Drift
Subtitle:  No alarm bells. No reset. Just continuation.
Closing line (key insight):  The danger is not drift. The danger is that the answer still looks correct.
The journey — how drift happens:
Train ticket to Crewe → Route, timing and options → Weather update → Wildlife sighting → Best café for breakfast
The two realities:
Human experience:  'I followed along naturally.' We trust the flow.
System reality:  No internal awareness of the change. Just continuation from the last turn.
The key warning:
Drift is normal. Valuable even. Risk comes from assuming it's still on track.
NOTE: Powerful chapter for executives. The business meeting equivalent: started talking about Q3 targets, ended up in a detailed conversation about office furniture, and the AI gave confident, coherent answers throughout. Nobody noticed the drift.

SLIDE 5 — Chapter 5: Compression: The Silent Transformation
Subtitle:  As the context window fills, older content is compressed.
Closing line:  It's not forgetfulness. It's the context window.
What you think is still there vs What Claude actually retains:
Exact details  → Abstraction (the gist, not the specifics)
Nuance and caveats  → Simplified summary
Quotes and numbers  → Approximation or omission
Earlier decisions  → High-level intent
Visual metaphor from the slide:
Sharp mountain peak (early turns) → Blurred mountain (older turns). The image sharpens the argument: what was crisp becomes soft. Not gone — just no longer precise.
NOTE: This chapter and 5.5 may be combined or run as paired chapters. Paul's call. But the distinction matters: Slide 5 is about what happens to the content. Slide 5.5 is about why — the architectural limit.

SLIDE 5.5 — Chapter 5.5: The Context Window Has a Limit
Subtitle:  As the conversation grows, earlier turns are summarised, regress, and eventually vanish.
Closing line:  The model does not forget like a human. It is the context window that runs out of room.
The conversation timeline — what Claude retains at each stage:
Turn 1–3 (earliest)  Exact details, exact wording
Turn n  Key points, specifics
Turn n+10  Main ideas, some detail
Turn n+20  Nuance and caveats are the first to go
Turn n+50  High-level summary
Turn n+75  Faint recollection
Turn n+100 (latest)  Vanished
What this means:
Earlier context becomes less precise.  
Nuance and caveats are the first to go.  
Decisions and agreements can become distorted.  
By the time it vanishes, it can no longer be corrected.  
NOTE: This is the chapter that lands hardest with executives who use AI for long strategic conversations. The decision made in Turn 3 is gone by Turn 75. The AI is still answering confidently. Nobody told you.

SLIDE 6 — Chapter 6: Fluency Creates False Confidence
Subtitle:  Because it sounds right, we assume it is right.
Sub-heading:  Claude produces fluent, coherent language even when the underlying context is incomplete, compressed or off track.
The central contrast:
What we experience (human interpretation):
It sounds confident.  So it must be correct.
It answered precisely.  So it must remember.
It knows my intent.  So it understands me.
It stayed on topic.  So it's following along.
It feels consistent and reliable.  So I can trust it.
What is actually happening (system reality):
It sounds confident.  → Probability, not certainty. Fluent completion, not knowledge.
It answered precisely.  → Compression fills the gaps. Details are approximated or omitted.
It knows my intent.  → Pattern recognition, not understanding. Statistical inference, not intention.
It stayed on topic.  → Continuity, not comprehension. It extends the path, not the purpose.
It feels consistent.  → Consistency by design. System prompts + preferences + patterns.
The Takeaway:
Fluency is not understanding. Coherence is not correctness. Always apply healthy verification. Especially when it matters most.
Executive Implication — AI literacy now includes:
1.  Question the output
2.  Check the context
3.  Verify the important
4.  Make humans accountable
NOTE: This is the most important chapter for the book's core argument. The clinical parallel is exact: a client who speaks with total conviction is not necessarily speaking truth. Fluency of delivery is not evidence of accuracy. Paul has thirty years of that lesson.

SLIDE 7 (deck slide 8) — Chapter 7: Context, Models, and What You Control
Subtitle:  Different access. Different limits. Different responsibilities.
User tiers and context windows:
Free  ~4K–32K tokens. Good for getting started. Limited access.
Pro  ~32K–200K tokens. Power user. Higher limits, priority access.
Team/Expert  ~200K–1M tokens. Advanced use. Your own environment.
Enterprise  Configurable (1M+ possible). Business grade.
What fills the window:
Your messages, Claude's responses, uploaded files, data, tables, code, system instructions, conversation history.
Enterprise advantage — your own LLM environment:
Private network, your LLM (hosted for you), your data stays yours. Data never leaves your environment. No training on your data. Custom models, fine-tuned for your domain. Compliance: GDPR, HIPAA, ISO, internal policies. Audit logs, monitoring and governance.
NOTE: This chapter is where Paul can introduce the Fuse Energy example directly — an ordinary consumer using AI to talk to a company's AI. The real-world gap made visible.

SLIDE 8 — Chapter 8: Behaviour Alignment
Core idea:  Humans must adapt their communication behaviour to AI operational reality.
The contrast:
Keep talking like a human:
—  Long, open-ended prompts
—  Implied context and vague goals
—  Story first, question later
—  Expecting memory and empathy
—  Optimised for conversation
Outcome:  Mediocre, inconsistent, drifting.
Speak the system's language:
—  Direct, goal-first prompts
—  Explicit context and constraints
—  Question first, then detail
—  No memory assumed
—  Optimised for clarity and precision
Outcome:  Better, faster, more reliable.
The argument:
It's not about being rude. It's about being effective.
NOTE: Paul's Fuse Energy post is the lived example for this chapter. He used his own AI to phrase a question to Fuse's AI in its dialect. That's behavioural alignment in practice, by someone who didn't know that's what it was called.

SLIDE 9 — Chapter 9: Sub-Personalities for AI Fluency
Subtitle:  Develop these modes to get better outcomes.
Closing line:  AI fluency is not one personality. It's a set of operating modes.
The six modes:
The Director  Sets clear goals. Defines success. Gives context first.
The Structurer  Organises information. Prioritises. Frames the context.
The Clarifier  Asks sharp questions. Seeks precision. Removes ambiguity.
The Challenger  Tests conclusions. Requests evidence. Checks logic and sources.
The Analyst  Breaks down output. Verifies. Checks logic and sources.
The Reflector  Learns from results. Refines approach. Improves over time.
The argument:
The suggestion is not personality fragmentation but intentional mode selection based on task and outcome. Different rooms require different stances. The same is true of AI interaction.
NOTE: Paul's clinical parallel: sub-personalities are a core tool in his practice (Russian Doll Therapy, psychosynthesis, parts work). This chapter lands in his native territory. He should write into this from personal authority, not from the slide alone.

SLIDE 10 — Chapter 10: Same Language or Different Game?
Subtitle:  You have a choice.
The tension:
Conversational human behaviour versus system-optimised interaction. Naturalness versus efficiency. Clarity versus precision. Emotional expectation versus operational effectiveness.
The argument:
Most people keep talking to AI the way they talk to humans. They get human-shaped responses that drift, compress, and lose fidelity over time. The people who adapt their communication to the system get better, faster, more reliable outcomes. The question is not which is more natural. It's which is more effective for what you need.
NOTE: This chapter sets up WaaS (Chapter 11) — if organisations adapt to system language at scale, they stop needing humans to translate. That's the commercial consequence.

SLIDE 11 — Chapter 11: WaaS — Worker as a Service
Subtitle:  Experience already shows the shift.
The frame:  SaaS: You buy software. People do the work. WaaS: You buy outcomes. AI does the work.
Closing line:  Different model. Different world.
Why people will choose AI:
—  Quicker — instant responses, 24/7
—  Simpler — no call queues, no transfers
—  Factually honest — no politics, no spin
—  Consistency at scale — same quality every time
—  Continuously improving — learns and gets better
Where humans remain preferred:
—  Empathy and nuance
—  Complex judgement
—  Trust and relationships
—  Sensitive situations
—  Creative situations
—  Creativity and originality
The argument:
Businesses will offer a choice. Most customers will choose AI for most things. The gap between what AI sounds like and what it is becomes an organisational risk, not just a personal one. When AI does the work, human accountability for verifying its outputs becomes critical.
NOTE: The WaaS image Paul supplied (robots at workstations vs humans) is the visual anchor for this chapter. The Fuse Energy example is the chapter in miniature: a consumer-facing AI agent doing the work of a customer service team.

SLIDE 12 — Chapter 12: Audio, Wearables and the Next Interface
Subtitle:  The conversation is leaving the screen.
What's changing:
Voice  Becomes the new default. Fast. Free. Hands-free.
Wearables  Provide always-on context. Location, activity, biometrics, environment.
AI  Becomes proactive, not just reactive. It anticipates needs and acts.
Why it matters:
Privacy  Sensitive data stays private.
Control  You set the rules, models and access.
Performance  Built for your use case and scale.
Know your limits. Choose your tier.
The closing argument for the chapter (Paul's framing):
Every slide in this book has assumed a screen. You type. It responds. You read. That assumption is ending. When the screen disappears, so does your ability to check. The same system. The same gap. But now there is no text to re-read, no pause before you respond, no visible output to question. The behaviours that will protect you depend entirely on which interface you're using.
The gap doesn't close when the screen disappears. It widens.
NOTE: This is the book's final chapter. It should close with the book's central line — Mind the Gap — reframed for the ambient AI world. Paul's voice here, not theory.

SECTION 3 — THE AUTHOR: THE LAYERS MAP
Source: Paul Roebuck — The Map of the Layers v4. The architecture beneath the practice. This is the author's identity, authority, and voice. Every chapter should be written from this foundation.

The single load-bearing description
Paul Roebuck is a working-class commercial operator turned psychotherapist, shaped by early loss, corporate pressure, cancer, and clinical practice. Across boardrooms, therapy rooms, AI work, grief theory, photography, and In-TEO, his consistent discipline is behavioural attention: listening to what people say, noticing what they do not, and helping identify what stops them changing.
The four public descriptions
Thin (6 words):  I pay attention to behaviour for a living.
Bio block (3 sentences):  For thirty years, personally, professionally in boardrooms and more lately as a psychotherapist, I've been asking one question: why do people do what they do — and what stops them doing it differently. I listen to what people say, and hear what they don't. I pay attention to behaviour for a living.
Thicker:  I'm a psychotherapist and coach. For thirty years I've been asking one question: why people do what they do, and what stops them doing it differently. I listen for what they don't say.
Thick (back cover):  I'm a psychotherapist and coach who has spent thirty years in boardrooms, consulting rooms, and more recently in AI rooms, asking why people do what they do and what stops them doing it differently. I listen for what they don't say. I pay attention to behaviour for a living. From that listening I've produced three originated clinical methods, a research-ready cycle of grief, and a piece of software for preparing people for the loss they cannot avoid. I lost my father at three. I survived mouth cancer at fifty. Both shaped the work. The voice consolidated post-cancer into the form that carries it now.
The seven floors
Floor 1 — Biological substrate:  The amygdala fires before consciousness arrives. The autopilot produces behaviour before the prefrontal cortex catches up. Insight alone doesn't change behaviour. The practice is partly about widening the gap that biology keeps trying to close.
Floor 2 — Personal substrate:  Originating wound: loss of father at age three. Not-good-enough. Cancer (2017) as second authoring event — clarifier and catalyst. Graduation poem closer (August 2008): I am good enough; I am ready to be the greatest I can be.
Floor 3 — Clinical floor:  Bowlby (attachment theory). Klein, Yalom (transference, projection). Assagioli, Ferrucci, Firman (psychosynthesis, sub-personalities). Bond (Behind the Masks). Originated methods: Russian Doll Therapy (2019), NGE-FOF Continuum (2022-2026), 12-Stage Cycle of Grief (May 2025).
Floor 4 — Philosophical priors:  Frankl: between stimulus and response there is a space. Covey: the way we see the problem is the problem. Buddha (Dhammapada): we are what we think. Berle: if opportunity doesn't knock, build a door. Sullivan: form ever follows function.
Floor 5 — Listening discipline:  Freud: evenly suspended attention. Bion: reverie, negative capability, without memory or desire. Reik: the third ear. Bollas: the unthought known. Rogers: empathic listening including what's not said. Kohut: selfobject hunger. Gestalt: figure and ground. Favoured channel: negative-space listening — what isn't said. The omission. The missing piece. The gap in the pattern. The same instrument that hears what a model didn't say.
Floor 6 — Three postures:  Reflective practitioner. Model originator. Founder (In-TEO, v46 prototype).
Floor 7 — Manifestation:  Therapy. Commerce. AI. Photography. Grief threading all four.
The two binding agents
Bowlby writing ethic:  Clarity. Precision. Honesty. Utility. No righteousness moralising. Write for the intelligent layperson. Concepts, not conclusions. Invite thought, not agreement.
Structural transposition:  The cognitive move that lets one instrument work in many rooms. Analogical reasoning. Isomorphism-spotting. The same capacity expressed across domains.
The key instrument for this book
Negative-space listening. What isn't said. The omission, the missing piece, the gap in the pattern. The same instrument that spotted commercial opportunity in unfilled markets and that hears what a model didn't say. This is the thread that connects Paul's thirty years of clinical work to the book's central argument. The gap is his native territory.

SECTION 4 — REAL WORLD EXAMPLES

The Fuse Energy / Plugs Example (Facebook post, 27 April 2026)
Paul's post to the Fuse Energy UK Customers Facebook group, titled: 'The Fuse AI: Plugs; making the AI agent work for you.'
Key quotes from the post:
—  I joined ChatGPT the week it launched in November 2022, and have Claude on paid.
—  So. Plugs. It speaks English but it's with its own dialect.
—  I had a detailed technical query about how my car, smart meter, car charger, and their respective apps managed off peak charging, schedules and various readings. There was a clear scheduling issue. So I wanted to ask Plugs. In its dialect.
—  I asked my AI to phrase the question. I sent it images of the apps and stuff. It compiled an incredibly detailed question and asked for a thorough reply. Which I got. Completely satisfied.
—  It's about specificity. Be precise. Use it conversationally. Chat to it like it's a WhatsApp. Explain and ask it to replay to you what you think you're asking.
—  It can't make decisions. It can and will answer any question you ask it. Literally.
—  It's just a dialect.
—  You've joined Fuse because of their tariff. You've joined Fuse AI because it's their help desk. By this time next year all your comms will be with AI Agents. Talk to them in their dialect.
The Top 10 Fuse Dialect help card Paul compiled from asking Plugs directly:
1.  Be the 'Where' and 'What' — mention your specific property or supply
2.  Ditch the Jargon — speak naturally, it translates
3.  Skip the ID Check — you're already logged in
4.  Ask for the 'How-To' — step-by-step app directions
5.  Describe, Don't Assume — it can't see your physical meter
6.  Facts Over Fiction — it will tell you rather than guess
7.  Stay in the Fuse Lane — expert on energy, not broadband
8.  Security First — big changes go to a human agent
9.  Emergency Protocol — gas leaks, use emergency numbers immediately
10.  The Human Bridge — if it feels too complex, ask for a human
NOTE: This example is the book's thesis in action. An ordinary consumer using one AI to talk to another AI's dialect. Behavioural alignment demonstrated without knowing that's what it was called. Use in Chapter 7 or Chapter 8.
The WaaS Frame
SaaS:  Software-as-a-Service. You buy software. People do the work.
WaaS:  Worker-as-a-Service. You buy outcomes. AI does the work.
Outcome delivered example:  Accuracy 99.6%, Cycle Time -68%, Cost to Serve -57%.
Closing line:  Different model. Different world.
NOTE: Use in Chapter 11. The commercial consequence of the interpretation gap at organisational scale.
The NGE-FOF Image
Visual: shape-sorting toy — triangle, cylinder, two blocks with mismatched holes. A hammer and mallet alongside. Caption: 'Are you the wrong shape for the world you're in?' Credit: Paul Roebuck — NGE-FOF Continuum.
The NGE-FOF Continuum: Not-Good-Enough to Fear-Of-Failure. Paul's originated clinical model, developed 2022-2026. Originally a binary, matured into a spectrum with range as the developmental goal.
NOTE: Potential use in the book's introduction or Chapter 1 — the human who is the wrong shape for the AI world they're now living inside. The gap between the person they are and the interaction style the system needs.

SECTION 5 — WRITING INSTRUCTIONS FOR CLAUDE

Voice
First person throughout. Paul's register: direct, exact, grounded, unsentimental. No waffle, platitudes, or inflated tone. The Bowlby ethic: clarity, no jargon, say what you mean, serve the reader. Short sentences. Short paragraphs. White space.
Not corporate. Not therapeutic. Not AI-generated smooth. Paul's voice is working-class in origin, clinically trained, commercially shaped, cancer-sharpened. It does not perform.
Structure of each chapter
Open:  A real scene. A moment. A conversation. Something that happened. Not a thesis statement.
Build:  The framework from the slide. Translated into prose. Theory following story.
Close:  A precise, single-sentence landing. Often a reframe. The chapter's gap named.
What to preserve from the slides
Every contrast table. Every closing line. Every key insight. These are the framework's bones. The prose builds around them, not over them.
What to add from The Layers
Paul's clinical authority. His thirty years. The therapy room parallels. The listening discipline. The wound that forged the practice. Use selectively and precisely — not as biography, as evidence.
What to avoid
—  Inflated language. No 'transformative', 'revolutionary', 'unprecedented'.
—  Technical jargon without translation.
—  Generic AI book framing — doom, boom, singularity.
—  Therapeutic language bleeding into the general argument.
—  Chapters that sound like the deck rather than like Paul.
Attribution (all published outputs)
Ideas and direction: Paul Roebuck — [platform/URL]
Words: Claude AI (Anthropic) | Images: [AI tool] | [Month Year]

End of Master Reference Document
Mind the Gap — Paul Roebuck — compiled May 2026