Turn ChatGPT Into a Chef: First Principles for Product Managers

Apr 16, 2026

9 min read

6 views

85%

Are you asking a tool to cook for you, or to teach you how to cook?

Every product manager has used quick AI outputs to generate PRDs, personas, and homepage layouts. Those outputs feel like a shortcut. They speed up churn, reduce writer's block, and create material you can iterate on. But shortcuts have a cost: if you rely on the recipes, you stop practicing the craft. If you never take products apart and recompose them from fundamentals, you will reproduce other peoples' assumptions and miss opportunities that only reveal themselves when you rebuild from first principles.

This article argues a single proposition: the real value of AI in product work is not in producing polished artifacts, but in amplifying first principles thinking. Treat AI as a generative chef: someone who knows ingredients and combinations, who can propose novel recipes when you first deconstruct the problem. That requires a different discipline than prompt copy-and-paste. Below I show a concrete framework to convert AI outputs into experiments, metrics, and new product intuition, with prompts and guardrails you can use today.

The setup: the practical power and the hidden trap

AI tools give product teams immediate abilities: brainstorm app concepts, draft PRDs, synthesize user feedback, produce UX copy, and compose go-to-market plans. In practice these are valuable. When you are stuck on wording for a difficult customer email, a model will save you five to twenty minutes and reduce friction.

But speed alone does not equal understanding. There are two common failure modes I see in teams that adopt AI too eagerly.

First, reasoning by analogy. Outputs often rely on patterns from existing products. That is helpful when you need a plausible baseline, but dangerous if you want a breakthrough. A recipe tells you what worked for many cooks; it rarely suggests the ingredient that would create a new cuisine.

Second, assumption drift. When models draft PRDs, personas, or feature lists, they implicitly bake in assumptions about customers, value, and distribution. Teams can skim these assumptions and move to design and build without testing them. If growth depends on distribution instead of product fit, or if revenue relies on an unjustified pricing assumption, the result is wasted time and missed insights.

A simple analogy clarifies the tradeoff: the difference between a chef and a cook. The cook follows recipes and produces reliable dishes. The chef understands ingredients at a molecular and cultural level and invents new recipes. In product management, AI excels at being a fast, tireless cook. Your job, if you want to build something distinctive, is to use AI to become the chef.

The tension: speed versus understanding, templates versus fundamentals

Here is the core tension to resolve: AI accelerates iteration, which should be liberating. But acceleration without a clear architecture of thought is acceleration toward convergence on the same local optima everyone else has found. That looks like shipping faster, while learning less.

First principles thinking offers a remedy. At its heart, first principles means decomposing a problem to propositions that cannot be deduced from anything simpler. It is an act of dismantling assumptions to test their truth. In product work this looks like asking: what are the minimal truths about this user, this behavior, this technical constraint, and this distribution channel? When you strip away borrowed recipes, what remains?

Socratic questioning gives the operational form: clarify assumptions, challenge them, look for evidence, consider alternatives, and examine implications. AI can support every step, but only if you prompt it to operate under that discipline rather than to produce plausible artifacts on demand.

Consider a concrete, common task: generating a user persona. A naive prompt produces a useful template, but that persona often blends stereotypes and generic needs. A first principles approach instead asks: what behaviors are invariant under context? What motivates action? What constraints make certain solutions impossible? Answering those questions reveals different personas and different experiments than a templated persona would suggest.

The paradox is this: speed without scrutiny amplifies error. Slowness to question becomes the bottleneck. First principles thinking is the tool that turns speed into leverage.

Synthesis: a practical playbook to use AI like a chef

Below is a repeatable framework you can use immediately. I call it PROBE: Prepare, Reduce, Observe, Build, Experiment. Each step shows how to combine first principles with AI capability.

Prepare: frame the true question

Start by writing one sentence that captures the hypothesis you truly want to test. Resist the urge to start with deliverables like PRDs or marketing plans. The one-sentence hypothesis forces clarity. Example: "Expecting mothers will pay for short daily meditations that reduce anxiety during the third trimester because they want 5 minute rituals they can do while feeding or working."

Then ask the AI to play two roles: skeptic and ally. Ask it to list the top five reasons the hypothesis might be false, and the top five ways to validate it quickly.

Sample prompt: "I have this one-sentence hypothesis: [insert]. List five concrete reasons it might be false with evidence I can collect in one week, and then list five minimum experiments to falsify or support it. Keep each item to one sentence."

Reduce: decompose to first principles

Strip the hypothesis down to elemental propositions. For each proposition ask: is it a fact, an assumption, or a conditional? Label them. Use AI to surface hidden assumptions by instructing it to 'play devil's advocate.'

Example decomposition for a meditation app might include: users value sessions shorter than 10 minutes; users prefer audio content over text; anxiety reduction can be measured via self-report after a week. Identify which of these are assumptions.

Sample prompt: "Decompose this hypothesis into its atomic propositions. For each proposition label it as fact, assumption, or conditional, and suggest one low-cost way to test it in two days."

Observe: design minimal experiments that prioritize learning

Turn each assumption into a rapid experiment. Resist building a full app. Use landing pages, storyboards, ads, or simple prototypes. Ask AI to draft the smallest possible experiment, including the single metric that would make you change course.

Example: a landing page that promises "5 minute meditations for expecting mothers" with an email capture and a pricing test using two different price points. The single metric could be the conversion rate from visitor to email signup.

Sample prompt: "For each assumption, suggest one minimal experiment and the one metric that would cause us to iterate. Provide a 3-step plan and an estimated cost and time."

Build: use AI to generate targeted artifacts for the experiment

Now use generative power, but constrain it. Ask the model to produce content that is explicitly meant to be tested. For example, draft two UX copy variants, one that emphasizes convenience and one that emphasizes evidence of effectiveness. The goal is falsification, not perfection.

Example prompts:

"Write two email subject lines to test click-through rate: one focused on time savings, one focused on anxiety reduction."
"Create a 90-second prototype script for a meditation aimed at 32-week pregnant women that includes an optional breathing cue at 40 seconds."

Experiment: collect evidence and update beliefs

Run experiments fast and record what they teach you. Use the AI to synthesize interview notes, user feedback, and quantitative metrics. But do not accept the synthesis as definitive. Pair AI summaries with a manual pass to catch nuance and coordinates of emotionality that models can miss.

Sample prompt: "Here is a transcript of a 20 minute interview. Summarize the top three insights and list any statements that contradict our core hypothesis. Put each insight as a testable statement."

Repeat the PROBE cycle until you have replaced assumptions with validated facts. The result is a product hypothesis that is falsified or strengthened by evidence, not just plausibility.

Concrete examples and templates you can use now

The value of this framework is practical. Below are concrete templates and a three hour plan you can adopt in a sprint.

Three hour rapid validation sprint

Hour 1: One-sentence hypothesis plus devil's advocate. Use the Prepare and Reduce steps to produce a list of atomic assumptions.
Hour 2: Design two minimal experiments and two variants of marketing copy or prototype screen. Keep experiments that can run on a landing page or via six interviews.
Hour 3: Create assets with AI, launch them to a small but appropriate audience, and draft an interview script.

Sample first-principles prompt template

"Hypothesis: [one sentence]. Decompose this into atomic propositions. For each, mark it as fact, assumption, or conditional and propose one two-day experiment to test it. Limit to six propositions."

Sample 'challenge assumptions' prompt

"Here are five assumptions we are making: [list]. For each, give three real-world counterexamples that would falsify it, and one supporting data point or source we should seek."

Sample 'UX copy for testing' prompt

"Write two variants of homepage hero copy to test with A/B. Variant A emphasizes convenience in one sentence. Variant B emphasizes measurable outcome in one sentence. Keep each under 120 characters."

Use these prompts to make AI the instrument of critical thinking rather than a source of canned conclusions.

Guardrails: when to trust AI outputs and when to distrust them

AI is excellent at pattern completion. Trust it for speed, iteration, and surface-level drafts. Distrust it for unseen contexts, novel mechanics, or when assumptions matter more than execution. Here are practical rules of thumb you can apply on every deliverable.

If the work is a hypothesis that needs validation, treat AI output as a source of test cases, not final answers.
If the output could acclimate your team to a false consensus, force a skeptical read. Ask someone to document the assumptions before acting.
Preserve human-in-the-loop synthesis. Use AI to summarize user feedback, but verify at least three representative quotes yourself.
Institutionalize a habit: every AI-generated PRD or persona must include a one-paragraph list of its core assumptions and the experiments planned to test them.

If you do not make assumptions explicit, AI will make them for you. Those assumptions will often be the softest, most dangerous ones.

Key Takeaways

Use AI to accelerate learning, not to skip it: convert outputs into falsifiable hypotheses and experiments.
Decompose before you generate: label propositions as fact, assumption, or conditional, and test the assumptions first.
Treat generative outputs as drafts for experiments: design minimal tests, collect data, then iterate.
Maintain a human-in-the-loop for emotional nuance and to catch AI blind spots: verify representative user quotes manually.
Institutionalize assumption-checking: every artifact must include a short plan to validate or falsify its key assumptions.

Conclusion: reframe the role you ask AI to play

If you adopt one habit today, make it this: train your prompts to ask the hard questions before you ask for deliverables. The greatest misuse of AI is to let it automate the craft that made you valuable in the first place: the ability to see what others do not, to question the obvious, and to design clean experiments that reveal reality.

Think of yourself as the chef and AI as your apprentice. You teach it the ingredients and the vision. It offers combinations and speeds up trial and error. But you decide which assumptions to smash, which experiments to run, and which recipes become your signature. That is where product advantage lives.

If you start designing prompts that demand first principles, you will stop shipping well-crafted copies of other people’s products. Instead you will build things that test new truths and, sometimes, change the game.