TL;DR
The V1 Framework is a five-step methodology for building with AI: Strip, Decompose, Constrain, Define Done, Instruct. It is built on the premise that the prompt is the smallest part of building good AI output. The real work is the thinking that precedes the prompt: stripping assumptions, decomposing the problem, defining constraints, and writing acceptance criteria. 90% thinking, 10% prompting. The framework works for prompt engineering, AI agent design, RAG systems, and AI strategy.
- The prompt is the specification. The thinking is the engineering.
- Strip first. Most AI failures are upstream of the prompt.
- A clear definition of done makes prompting trivial.
In this article
Why the V1 Framework exists
The V1 Framework came out of a practical problem. I was shipping production AI inside CreativeOS, advising on AI strategy at consumer brands through Automatic, and watching teams everywhere produce wildly different output quality with the same models. The variable was not the model. It was the thinking.
Teams that thought clearly produced great AI output with average prompts. Teams that thought sloppily produced mediocre AI output with elaborate prompts. The pattern was obvious enough to name.
The V1 Framework names it. It is the operating methodology I use to build production AI, and it is the methodology I teach when I am brought in to fix AI programs that are stalled.
It is called V1 because the work of building with AI is the work of getting to a clear v1. Not the work of optimizing v8.
The 90% thinking, 10% prompting principle
The shortest version of the framework is this: clarity is the prompt. Everything else is typing.
Most teams reverse the ratio. They spend 10% of their effort on the thinking, then spend 90% writing elaborate prompts to compensate for the missing thinking. The output reflects that imbalance. It looks clever and lands wrong.
The right ratio is 90% thinking, 10% prompting. That means:
- 90% of your time is spent stripping assumptions, decomposing the problem, defining constraints, and writing acceptance criteria.
- 10% of your time is spent writing the prompt that converts the thinking into instructions.
Teams that adopt this ratio ship faster, debug faster, and produce more reliable AI output. The reason is simple: if the thinking is clean, the prompt almost writes itself. If the thinking is dirty, no prompt will save it.
Clarity is the prompt. Everything else is typing.
The five steps
The V1 Framework has five steps. They are sequential, not parallel. Skipping any of them produces predictable failure modes.
Step 1: Strip
Strip removes the assumptions inherited from how the work was done before AI. What does the user actually want? What is the actual job to be done? What inputs are actually available?
Most AI failures live here. Teams describe the existing workflow to the AI instead of describing the desired outcome. The AI dutifully reproduces a broken process at higher speed. Strip is the work of pulling the assumptions off the problem until only the actual job remains.
Step 2: Decompose
Decompose breaks the stripped problem into its truth-level components. What is the smallest unit of work? What are the dependencies between units? What can be parallelized and what is sequential?
Decomposition is the step that turns a vague "build me an AI agent that handles customer support" into a list of discrete tasks the AI can actually do. Without decomposition, you are asking the AI to do everything at once. With decomposition, you are asking it to do one thing at a time, the right number of times.
Step 3: Constrain
Constrain defines the real boundaries of the AI's authority. What can it decide? What must it escalate? What inputs are valid? What outputs are valid? What is the failure mode?
Constraint is the step that most teams skip because it feels limiting. It is actually liberating. A well-constrained AI is fast and reliable. An unconstrained AI is slow and hallucinatory. The constraint is the operating spec.
Step 4: Define Done
Define Done writes the acceptance criteria for the AI's output. What does a good output look like? How will we know? Who decides?
Most AI evaluation problems are actually undefined-done problems. The team has not written down what success looks like, so the AI cannot aim at it and the humans cannot judge it. Define Done is the contract between the thinking and the output.
Step 5: Instruct
Instruct is the prompt. It is the last step, not the first. If the previous four steps are clean, the prompt is short and obvious. If the previous four steps are dirty, the prompt is long and clever and the output still misses.
The prompt is the specification. It carries the constraints, the definition of done, and the decomposed task into the model. The art of the prompt is not artistic. It is the work of rendering clear thinking into clear instructions.
Where the V1 Framework applies
The V1 Framework was developed for prompt engineering, but it applies anywhere AI is used to produce work:
- Prompt engineering. The original use case. Writing prompts that produce reliable output.
- AI agent design. Designing multi-step agent workflows. Each step is a V1 in miniature.
- RAG architecture. Defining what gets retrieved, what gets ignored, and how the model uses context. See RAG architecture for consumer brands.
- AI strategy. Applying the same thinking discipline at the program level. What is the actual job the AI program is doing? Strip the assumptions. Decompose the workflow. Define done.
- Evaluation harness design. Define Done is the same artifact whether you are scoring one prompt or running a continuous eval against production.
The principle scales. The work of clear thinking does not change as the scope changes.
How to learn the framework
The V1 Framework is learnable in a week of practice and ownable in a quarter of work. The fastest way to learn it is to apply it to a real problem you are currently stuck on:
- Pick a real AI workflow you are building or trying to fix.
- Read the five deep-dive articles in this series.
- Apply each step to your workflow in order, in writing.
- Write the prompt last. Compare the result to your previous version.
- Run the eval set you should have built in Step 4.
The framework is not a credential. It is a habit. The teams that adopt it stop producing demoware and start producing systems. The teams that ignore it stay in demo purgatory. For more on the distinction, see Production AI vs AI Demos.
For the deep-dives, start with Step 1: Strip and move through in order.
The bottom line
The V1 Framework is the methodology behind every production AI system I have shipped. It is 90% thinking and 10% prompting because that is the ratio that produces reliable output. Strip the assumptions, decompose the problem, define the constraints, write the acceptance criteria, then write the prompt.
The prompt is the last 10% of the work. The first 90% is what separates the demo from the system. For the broader program context, see the AI transformation playbook for consumer brands.
FAQ
What is the V1 Framework?
The V1 Framework is a five-step methodology for building with AI: Strip, Decompose, Constrain, Define Done, Instruct. It treats the prompt as the final 10% of the work, with 90% of the value coming from the thinking that precedes the prompt.
Who created the V1 Framework?
The V1 Framework was developed by Nicholas Harris, President at CreativeOS and Founder at Automatic, based on shipping production AI systems for consumer brands. It is the methodology behind production LLM deployments serving 25,000+ brands.
Is the V1 Framework just prompt engineering?
No. Prompt engineering is the last step of the V1 Framework. The framework is the broader thinking discipline that produces good prompts as a byproduct. Teams that focus only on prompting without doing the upstream thinking produce unreliable output.
How is the V1 Framework different from chain-of-thought prompting?
Chain-of-thought is a prompting technique. The V1 Framework is a methodology that includes prompting as one of five steps. Chain-of-thought lives inside Step 5: Instruct. The framework dictates what thinking the model should be doing; chain-of-thought is one tactic for getting it to do that thinking.
Can the V1 Framework be used with any LLM?
Yes. The V1 Framework is model-agnostic. It works with GPT, Claude, Gemini, Llama, and any other LLM. The framework is about the thinking that precedes the prompt; the prompt itself is rendered into whatever model you are using. For model selection, see how to pick the right LLM.
How long does it take to learn the V1 Framework?
The framework is learnable in a week of focused practice on a real problem. It is ownable in a quarter, meaning consistent application across multiple workflows. The fastest path is to apply it immediately to an AI workflow you are currently building or trying to fix.