Teams that succeed with generated forms rarely treat generation as a single button press. They run a repeatable loop. The teams that struggle treat the first draft as the deliverable, ship it, and discover the problems through their respondents. The difference is process, not tooling.
This article names that process: the Draft, Critique, Refine model. It is deliberately small, three stages plus a decision rule for when to loop back. The goal is something you can hold in your head and apply every time, regardless of which builder you use.
We define each stage by its input, its output, and its exit condition. A stage is not done until its exit condition is met. Exit conditions are what separate a real process from a ritual. Without them, teams drift: they declare a draft finished because they are tired of looking at it, or they critique vaguely and call it review. Naming a concrete exit for each stage forces honesty about whether the work is actually done.
Why a model beats ad hoc generation
Ad hoc generation produces inconsistent results because each form starts from a slightly different prompt and gets a different amount of review. A named model standardizes both the input and the scrutiny, which is what the agency in our intake case study discovered when they formalized their process.
The core idea
Generation is cheap; judgment is the constraint. The model front-loads judgment into a structured critique stage so that problems surface before launch rather than after. This inverts the instinct most people bring to these tools. The natural impulse is to spend effort getting the draft perfect, then ship. The model says the opposite: spend little effort on the draft, because it is disposable, and concentrate your effort on critique, where judgment actually lives. Once you accept that generation is the cheap part, the whole process reorganizes around scrutiny.
Stage one: Draft
The first stage produces a complete first version from a structured prompt.
Input and output
The input is a prompt naming the goal, the audience, the decisions each answer informs, and any hard constraints. The output is a full draft form: questions, types, logic, and validation.
Exit condition
You exit Draft when the form is structurally complete, even if rough. Do not polish here. Polishing a draft you are about to critique wastes effort. The temptation to perfect wording in this stage is strong and counterproductive, because the next stage may cut half the questions you just labored over. Aim for completeness, not quality. A complete rough draft is a better input to critique than a beautiful partial one.
Stage two: Critique
This is the stage most teams skip, and skipping it is why their forms fail.
Self-critique then human critique
First, ask the model to critique its own draft against specific criteria: neutrality, scope, logic integrity, and privacy. Then a human reviews, because the model cannot judge domain nuance or organizational context.
What to critique against
Critique against the same passes in A Pre-Launch Review List Before Shipping a Generated Survey: structure, language, logic, data, and live behavior. The checklist is the operational form of this stage.
Exit condition
You exit Critique with a written list of specific changes. If the list is empty, you did not critique hard enough. The written list is non-negotiable, because critique held only in your head blurs into editing and you lose track of what you found. Writing the changes down also lets you see patterns: if three of your five notes are about the same kind of problem, your prompt has a systematic flaw worth fixing at the source rather than patching question by question.
Stage three: Refine
Refine applies the critique, then re-tests.
Targeted edits, not regeneration
Apply the specific changes from the critique. Resist regenerating the whole form, which discards the structure you already validated and reintroduces solved problems.
Exit condition
You exit Refine when every critique item is resolved and the live test passes. Then, and only then, you ship.
The loop-back rule
The stages are not strictly linear. Here is the rule for when to go backward.
When to return to Draft
If Critique reveals a structural problem, a missing whole section, or a fundamentally wrong approach, return to Draft. Small wording or logic issues do not warrant this; they are handled in Refine.
Knowing the difference
Structural problems are about what the form is. Refinement problems are about how a part is worded or wired. Misclassifying the two is the most common process error, and it inflates effort, a cost we examine in Counting the Dollars Behind Automated Questionnaire Drafting.
Why teams skip Critique, and how to stop
Naming the stage is easy; protecting it is hard. Understanding why it gets skipped helps you defend it.
The pressures that erode review
Critique gets cut for three reasons: the draft looks finished so review feels redundant, deadlines compress, and critique is less satisfying than producing something new. All three are predictable, which means you can plan against them. Build the critique step into your definition of done so it is not optional, the way the agency in our intake case study made review a required part of every form.
Making critique structural
The most reliable fix is to attach critique to an artifact, the written change list, and a tool, the pre-launch review list. When critique produces a tangible output that someone expects to see, it stops being the step that quietly disappears under pressure.
Applying the model in practice
For a simple internal form, the whole loop takes under thirty minutes. For a high-stakes survey, Critique may take longer than Draft and Refine combined, which is correct. The model scales its rigor to the stakes.
Tuning the loop to context
The same three stages run on every form, but the time you spend in each shifts with the stakes. A throwaway internal poll spends almost all its time in Draft and a minute in Critique. A regulated public survey inverts that ratio entirely. The model does not change; your allocation of attention within it does. Learning to read how much rigor a given form deserves is itself a skill, and it overlaps with the judgment in When Hand-Built Beats Generated, and the Reverse.
A worked pass through the model
Walking the model once on a concrete form shows how the stages and exit conditions behave together.
Draft in action
Suppose you need a customer cancellation survey. The prompt names the goal, understanding why customers leave, the audience, departing customers, and the decision each answer informs. The model returns ten questions. You stop the moment the form is structurally complete and resist polishing, meeting the Draft exit.
Critique in action
The self-critique pass flags two leading questions that assume dissatisfaction, and your human review notes a missing question about whether the customer might return. You write these down as a three-item change list. The list is non-empty, so the Critique exit is met honestly, and the missing-question note signals a structural gap worth a brief loop back to Draft.
Refine in action
You apply the three changes, add the return question, reword the two leading ones, and run a live submission. Every item resolves and the test passes, meeting the Refine exit. Only now does the survey ship. The whole pass took twenty minutes, and the structure mirrors the operational steps in A Pre-Launch Review List Before Shipping a Generated Survey.
Frequently Asked Questions
Why separate Critique from Refine?
Mixing them means you edit as you read, which causes you to miss systemic problems while fixing local ones. Producing a written change list first forces you to see the whole picture before acting.
Should the model critique its own work?
Yes, as a first pass. Self-critique reliably surfaces leading questions and scope problems. But it cannot replace human review of domain nuance, so always follow it with a person.
When do I loop back to Draft versus just refining?
Loop back to Draft only for structural problems: a missing section, a wrong approach, or a fundamentally misframed goal. Wording and logic fixes stay in Refine.
Does this model work for any form builder?
Yes. The model is tool-agnostic. It governs how you think about generation and review, not which software produces the draft.
How much time does the full loop take?
For a simple internal form, under thirty minutes. For a high-stakes public survey, Critique alone can exceed the other two stages, and that is appropriate.
Key Takeaways
- Generation is cheap and judgment is the constraint, so the model front-loads structured critique.
- Draft produces a complete but rough form; do not polish what you are about to critique.
- Critique is the stage most teams skip, and skipping it is why their forms fail in the field.
- Refine applies targeted edits rather than regenerating and discarding validated structure.
- Loop back to Draft only for structural problems, never for local wording or logic fixes.