The Framework-First Method
Build AI that performs — no matter which model you use.
In the video I explain what this guide will do for you. Short version: for the past 2 years we've been building AI-first companies, and almost every failed AI build we've seen fails for the same reason — and it's not the model. This page gives you the framework we use on every build, plus two prompts you can copy and use today.
Want your business prepared for AI? → The AI Business BlueprintEveryone blames the model. The model is almost never the problem.
When an AI agent or automation underperforms, the first instinct is always the same: "maybe I need a better model." So people upgrade, pay more, switch providers — and get the same mediocre results, just more expensive.
Here's what we've learned building AI systems every single day for two years: when the framework around the AI is right, almost any model performs. When the framework is wrong, no model can save you. The cheapest model inside a great framework beats the most expensive model inside a messy one.
And there's a simple reason people get this wrong.
Humans think in tools. Great AI builders think in data flow.
Watch how most people start an AI project: "Should I use ChatGPT or Claude? Zapier or n8n? Which agent platform?" Tools first. That's the human instinct — and it's backwards.
Watch how an AI coding agent designs a system: it never starts with tools. It starts with how the data flows — what comes in, what has to be understood, what gets decided, what happens next, and how you know it worked. The tools get picked last, to serve the flow.
Every strong AI system we've ever built follows the same flow:
To design that flow, you answer six questions — before you touch any tool, any model, any platform. These six questions are the framework.
Framework first. Model last.
The 6 questions of every AI build that works
We'll use one running example so this stays concrete: a service business that wants AI to handle its incoming leads.
What does "it worked" look like, in one sentence?
Not "handle my leads with AI." That's a wish, not an outcome. An outcome is specific enough that you could check it at the end of the week.
What information enters the system, from where, in what shape?
Every input you forget here becomes a bug later. Where do leads come from — a form, DMs, email, a phone call transcript? What fields arrive? What's often missing or messy?
What judgments does the system have to make — and by what rules?
This is the question almost everyone skips, and it's the one that separates a real AI system from an expensive toy. Decisions are where value is created or destroyed: who is qualified? What tone to reply in? When to escalate to a human? If you can't state the rule, the AI is guessing — and you're paying for its guesses.
What does the system actually do once it decides?
Send the email. Book the call. Update the CRM. Notify you on Slack. List every action explicitly — and which tool owns it. This is the FIRST moment tools even enter the conversation. Notice how late that is.
How does the system check its own work before it acts?
This is what separates systems you can trust from systems you babysit. Before the AI sends, books, or updates — what does it verify? Did the email address parse correctly? Does the reply mention the right service? Is it about to promise something you don't offer?
What numbers tell you it's working — or quietly failing?
A system without KPIs doesn't fail loudly — it fails silently, for weeks. Pick 2–4 numbers you'll actually look at.
Answer the six questions, and the model stops mattering so much.
When decisions have clear rules, verifications catch mistakes, and every step knows its job — the AI inside the system has a small, well-defined task instead of a vague giant one. Small, well-defined tasks are exactly what cheap models do brilliantly.
That's the secret behind "10x better AI": it was never about buying a smarter brain. It's about building a system where the brain can't get lost. Framework first. Model last.
The 4 multipliers — how a cheap model performs like a frontier one
The six questions design the system. These four multipliers are how you squeeze frontier-level performance out of whatever model sits inside it. This is where most builders leave 80% of the performance on the table.
At every decision point, put the rules and real examples in front of the model.
A model can only decide as well as what it can see. Your pricing rules, your definition of a qualified lead, two real examples of a great reply — give them to the model at the moment it decides, in the prompt or from your documents. A cheap model reading your rules beats a frontier model guessing them. This is the UNDERSTAND stage done right — and it's why documented businesses get 10x more from AI than undocumented ones.
Never hand AI one giant vague task. Split the flow into small steps, each with a single job.
"Read this lead, figure out if they're good, write a reply, and update the CRM" is four jobs pretending to be one — and that's where every model, cheap or expensive, starts failing. Split it: one step classifies, one step drafts, one step updates. Small, defined tasks are exactly where cheap models match expensive ones. Giant vague tasks are where they all break.
Two or three real examples beat paragraphs of instructions.
Instead of describing your tone for a page, show the model 2–3 real pairs of "this message came in → this is the perfect reply." Models copy patterns far better than they follow essays. Whenever a step underperforms, your first fix should be adding an example — not rewriting the instructions, and not upgrading the model.
Define what happens when the AI isn't sure. Never force an answer.
A model forced to always act will confidently do the wrong thing. A model with a defined escape hatch — "not sure → ask one question, escalate, or hold" — makes dramatically fewer costly mistakes. The way out isn't a weakness in your system; it's the feature that makes everything else trustworthy.
And the habit that compounds it all — close the loop: every time the system gets something wrong, don't just fix the output. Write the rule that would have prevented it, and add that rule to the system's knowledge. Builds that turn errors into rules get better every week. Builds that don't, quietly decay.
One honest note: will this literally turn a small model into a frontier one? No — raw capability is real. But most business tasks don't need raw genius; they need a defined job done reliably. On defined jobs, these multipliers close most of the gap — at a fraction of the cost.
Now — two prompts that put all of this to work for you today. Copy them into Claude, ChatGPT, or whatever you build with.
Fix an AI build that's underperforming
Have an automation, agent, or AI workflow that disappoints? Paste this, describe your build, and let the AI audit it against the framework.
I have an AI system that is underperforming, and I want you to audit it using the Framework-First Method before suggesting ANY tool or model changes. Here is my build: [DESCRIBE YOUR SYSTEM: what it's supposed to do, what tools/models it uses, what goes wrong or disappoints you] Audit it by answering these six questions about my system as it exists today: 1. OUTCOME — What outcome is this system actually built to achieve? Is it specific and checkable, or vague? Restate it in one measurable sentence. 2. INPUTS — What inputs does it receive, and where is the input messy, incomplete, or unstructured in ways that cause failures downstream? 3. DECISIONS — List every judgment the system makes. For each: is there an explicit rule, or is the AI guessing? Mark every unstated rule — this is the most likely source of my problems. 4. ACTIONS — List every action it takes. Are any actions happening without a decision that justifies them? 5. VERIFICATIONS — What does it check before acting? List every place it acts WITHOUT verifying first — each one is a silent failure waiting to happen. 6. KPIs — What numbers would reveal whether it's working? Which of them am I currently blind to? Then audit the four multipliers: - KNOWLEDGE: at each decision point, does the model have my actual rules, definitions, and real examples in front of it — or is it guessing from general knowledge? - TASK SIZE: is any single step doing more than one job? List the steps that should be split. - EXAMPLES: where would 2–3 real input→output examples outperform the current instructions? - ESCAPE HATCH: where is the AI forced to act even when unsure? Add a "not sure → ask, escalate, or hold" path. Finally give me: - The 3 biggest gaps across everything above, ranked by impact - The specific fix for each, in plain language - Only AFTER all of that: whether a different model or tool would even matter, and why Do not recommend upgrading models or switching platforms until you've shown the framework gaps first.
// Works in Claude, ChatGPT, Gemini — any capable model. That's the point.
Build new — the right way, from the start
Starting something new? This prompt forces the framework before a single tool gets picked.
I want to design a new AI system using the Framework-First Method. Do NOT suggest any tools, platforms, or models until step 3. What I want to build: [DESCRIBE IN PLAIN LANGUAGE: the task you want AI to handle, and the business context] STEP 1 — Interview me. Ask me the six framework questions ONE AT A TIME, waiting for my answer before the next: 1. What outcome? (push me until it's one specific, checkable sentence) 2. What inputs? (sources, fields, what's often messy or missing) 3. What decisions? (every judgment the system must make — push me to state an explicit rule for each one; flag any decision where I can't state the rule, because that's where it will fail) 4. What actions? (everything it does after deciding) 5. What verifications? (what it must check BEFORE each action; every unverified action is a silent failure) 6. What KPIs? (2–4 numbers that prove it's working) STEP 2 — Play it back. Show me the full system as a flow: INPUT → UNDERSTAND → DECIDE → EXECUTE → VERIFY, with my answers mapped onto each stage. Flag any gaps or contradictions in my answers. Then apply the four multipliers: show me exactly where to inject my rules and 2–3 real examples at each decision point (knowledge), split any step that's doing more than one job (one step, one job), and add a "not sure → ask, escalate, or hold" path anywhere the system would otherwise be forced to act (the way out). STEP 3 — Only now: recommend the simplest toolset and the CHEAPEST model that can run this flow reliably, and explain why the framework makes a more expensive model unnecessary. If a task genuinely needs a stronger model, isolate exactly which step and why. Rule for this whole conversation: the framework decides the tools. Never the other way around.
// Tip: your answer to question 3 is where the money is. Spend the most time there.
The framework has one requirement most businesses can't meet yet.
Look at question 3 again — what decisions, by what rules? To answer it, you need to know how your business actually decides things: who's a good customer, how delivery really works, what "done right" means. For most owners, all of that lives in one place: your head.
That's the real reason most AI projects stall. Not the model. Not the tools. The business was never documented well enough for AI to run any part of it.
That's exactly what we fix. The AI Business Blueprint is 30 days where we work for you — mapping how your business gets customers, delivers, and keeps them, and turning it into the complete, AI-ready foundation: your playbook, your SOPs, your data files, your build path. Everything this framework needs to actually work in your company.
Get your business AI-ready → The AI Business Blueprint