The Amplifier

The Model

The figures from the book

Every figure in the book is a loose pencil sketch. Desk-resolution versions will be added here as the final art is completed; until then, each is described so you know what to look for.

In brief

Ten figures appear in the book, each first shown in the chapter where its idea is taught. The two readers return to most are Figure 7.1 (the eight steps with their loop-backs) and Figure 7.2 (the Decision Filter as a napkin decision tree).

The five reference figures

These are drawn for the site rather than taken from the book, and they are the ones worth keeping open while you work. The pencil figures from the chapters are listed underneath, and will appear here as the final art is finished.

The Amplifier Framework: eight steps, and who acts at each one Eight numbered steps. Steps one to three are done by you alone and are what keep you from becoming Dependent. Steps four to eight bring AI in and are what keep you from becoming an Avoider. At every step you are the one deciding; AI contributes at steps two, four, five and six. YOU AI The Amplifier Framework Eight steps. At every one of them, you are the one who decides. 1 Define the End State What success looks like, written before anything else. 2 Map the Flow You map it. AI reviews it. You decide what changes. 3 Assign Tasks Each task through the Decision Filter: yours or the machine's. 4 Select the Tool Pick by category, not by habit. 5 Prepare and Direct Full context, expected output, criteria set in advance. 6 Execute You do your tasks. You direct AI on its own. 7 Evaluate Against your criteria — not against how polished it sounds. 8 Integrate and Stress-Test Every part can be right and the whole still fail. Keeps you from Dependent Keeps you from Avoider you act AI contributes AI has no part
The eight steps, and who acts at each one.
The three types, compared A comparison of the Dependent, the Avoider and the Amplifier across five things: how they treat steps one to three, what they do with output, the state of their judgement, how they respond when something breaks, and whether they can own the result. The three types Almost nobody is purely one. The question is which is your default. Dependent Outsources the thinking. Feels productive; becomes hollow. Avoider Refuses to engage. Feels principled; falls behind. Amplifier Thinks clearly, then uses AI inside that clarity. Steps 1–3 Skips them. Goes to the tool first. Refuses them along with everything else. Does them alone, before any prompt. The output Accepts it because it sounds polished. Never produces one this way. Judges it against criteria written in advance. Judgement Cannot evaluate — never built the muscle. Intact, but applied to less and less. Built and exercised every cycle. When it breaks Patches whatever the tool returns. Starts again by hand. Knows a quick fix from a root cause. Owning it Cannot explain or defend it. Owns work that arrives too late. Can explain it, defend it, build on it. Colour is a label here, never the meaning — the name is always beside it.
The three types across the five things that separate them.
The Decision Filter: four questions that assign a task Four questions to run a task through before deciding whether you do it or AI does it. Whether the context lives only in your head, whether the output is verifiably right or wrong, whether the task is generating options or choosing between them, and what being wrong would cost. If the answer is still unclear, the task is not broken down far enough. The Decision Filter Run every task through these four before you assign it. 1 Does the task need context that exists only in your head? If you cannot write down everything someone else would need, it is yours. if yes → you 2 Does it have a clearly verifiable output? If “good” is subjective rather than checkable, keep a human in the loop. if yes → AI 3 Is it generating options, or choosing between them? Generating is work a machine does well. Choosing is the part you own. generate → AI 4 What does being wrong cost? Low cost, let AI draft and review it anyway. High cost, do it or verify it yourself. if low → AI Still cannot answer? Then the task is not broken down far enough. Split it and ask the four questions again. From The Amplifier, Step 3. Vivek sketches this on a napkin.
The Decision Filter: four questions that assign a task.
Tool Selection Logic: five characteristics that pick the tool Five things to weigh when choosing which AI tool to use: the nature of the input and output, how much context it must hold, whether the work is one answer or a conversation, how accurate it has to be, and how often you will do it. Choosing the tool Five characteristics. Using one tool for everything is a screwdriver for every repair. Input and output What goes in, what must come out? Text to text is a language model. Image to text is vision. Text to image is a generator. Data to pattern is analytics. Context How much does it need to hold at once? Long documents need a large context window. Anything touching your internal data needs integration, not copy-paste. Shape of the work One answer, or a conversation? Iterative work wants a chat interface. One-shot jobs are usually better in a specialised tool. Accuracy What happens if it is confidently wrong? Anything factual needs web access or a verified source. Brainstorming does not. Frequency Once, or every week? Recurring work is worth setting up templates for. A one-off uses whatever is already open. Use AI to find AI Give a general model your five answers and ask it to shortlist three tools against them. Then you choose. From The Amplifier, Step 4. Tool names change every six months; these five do not.
Choosing the tool: five characteristics, no brand names.
Iteration: telling a small fix from a structural problem One test decides how to iterate: can you fix the thing without affecting anything else? If yes it is a small fix and you correct it where you stand. If fixing it breaks something else it is structural, and four symptoms each send you back to a specific step: a budget that will not add up and a plan that conflicts with itself go back to step two, poor AI output across a whole section goes back to step five or four, and work that is correct but does not feel right goes back to step one. Small fix, or structural problem? Iteration is not failure. It is the framework working as designed. Can you fix it without affecting anything else? yes no Small fix Local, self-contained, breaks nothing else. Correct it where you stand and carry on. LOOKS LIKE A temple closed on the day you planned to visit A flight that lands later than you assumed A restaurant shut on Mondays No buffer time between two bookings Structural problem Fixing one thing breaks another. Stop patching and go back to where it actually went wrong. SYMPTOM AND WHERE IT GOES BACK TO The budget will not add up Step 2 The plan conflicts with itself Step 2 AI output is poor across a whole section Step 5 or 4 Everything is correct but it does not feel right Step 1 From The Amplifier, after Steps 7 and 8.
Small fix or structural problem, and where each one sends you.

The figures in the book

FigureWhat it showsFirst appearsDownload
Figure 4.1The AI Landscape

A hand-drawn map of the major categories of AI tools — conversational text, coding, image, video, voice and transcription, agents — with example brands small beneath each. The emphasis is on the categories; the names are almost incidental.

Chapter 4Final art pending
Figure 4.2The Three Types

Three figures side by side: the Dependent slumped with the AI doing the work; the Avoider with back turned and arms crossed; the Amplifier upright, one hand on the tool, one on their own work. No arrows between them — they are choices, not stages.

Chapter 4Final art pending
Figure 5.1The Dependent's Trajectory

Two lines on a time axis. 'Visible output / praise' climbs steeply for a year or two, plateaus, then dips. 'Underlying capability' is flat from the start. The growing gap is shaded and labelled 'the trap'.

Chapter 5Final art pending
Figure 6.1The Asymmetry of the Two Traps

The Dependent: a single figure with arrows of harm curving inward. The Avoider: a figure surrounded by smaller figures with arrows radiating outward. Captions: 'Hollows the self.' / 'Imposes on others.'

Chapter 6Final art pending
Figure 7.1The Eight Steps of the Amplifier Framework

A horizontal flow of the eight steps, Step 2 showing its three sub-phases, with loop-back arrows from Steps 7 and 8. The foundational visual readers return to.

Chapter 7Final art pending
Figure 7.2The Four-Question Decision Filter

A napkin-sketch decision tree: a task enters, four questions in sequence, three landing zones — 'Human task', 'AI task', 'Decompose further'.

Chapter 7Final art pending
Figure 7.3Iteration: Small Fixes vs Structural Problems

Two columns. Small fixes as small marks on a finished plan with arrows to a quick fix. Structural problems as cracks running through the whole, with arrows curving back to Steps 1, 2, 3 and 5.

Chapter 7Final art pending
Figure 8.1Tracing the Structural Problem Back to Step 1

A first pass ending in 'Generic. Doesn't fit.', a curved arrow to Step 1, a second pass with a fuller end state and three lightly redone tasks, ending in 'Lands.' The redo was small; the difference was disproportionate.

Chapter 8Final art pending
Figure 10.1The Amplifier Stack

Three layers: domain expertise (largest, 'years to build, irreplaceable'), AI fluency ('months if the foundation is there'), human skills ('the differentiator, the lifelong layer'). Below: 'Your unique value'.

Chapter 10Final art pending
Figure 11.1The Framework, Applied to a Festival

The eight steps down the left; Kavya's festival version down the right in her hand — 'the girl at the gate', forty tasks sequenced, 'mine' versus 'the intern'. The two columns are the same shape. Not one word of the framework changed.

Chapter 11Final art pending

Printable versions of the eight steps and the Decision Filter are already available on the downloads page as the Quick Reference Card.

Continue

The FrameworkThe ideas behind Figures 7.1–7.3.DownloadsThe reference card and worksheets.