Most people refine AI output by instinct: read the draft, feel that something is off, type a vague correction, repeat. It works sometimes, fails unpredictably, and is impossible to teach. What teams need is a structure—a small set of named stages that turn refinement from a feel into a procedure anyone can run.
The Draft-Diagnose-Constrain method is that structure. It breaks every refinement loop into three distinct stages, each with a single job. Draft produces raw material. Diagnose identifies precisely what is wrong. Constrain fixes it with a targeted instruction. The discipline is in keeping the stages separate—most failed loops collapse diagnose and constrain into a single vague nudge, which is why they spiral.
This article defines each stage, the rule for advancing between them, and when to apply or skip parts of the model. It is the structural backbone behind the practical checklist for running refinement loops.
Why bother naming the stages at all? Because naming is what makes a practice teachable, measurable, and reproducible. An unnamed habit lives in one person's instincts and dies when they leave. A named method with explicit stages can be written down, handed to a new teammate, audited when it fails, and improved deliberately. The names are not decoration—they are the difference between a personal knack and a team capability.
Stage One: Draft
What It Does
Draft is the first generation. Its job is to produce raw material fast, not to be perfect. Treating the first output as a starting point rather than a verdict is the mindset shift the whole method depends on.
How to Run It Well
A strong draft prompt front-loads everything you already know you need: audience, format, length, hard constraints, and a reference example if you have one. The better the draft prompt, the fewer loops you need—but you should still expect at least one.
When You're Done
Draft is finished the moment you have something concrete to react to. Do not refine inside this stage; resist the urge to nudge before you have diagnosed.
Stage Two: Diagnose
What It Does
Diagnose is where you name precisely what is wrong with the current output. This is the stage people skip, and skipping it is the single biggest cause of spiraling loops.
How to Run It Well
Articulate the defect in concrete terms before you write any instruction. "The third section repeats the second" is a diagnosis. "Make it flow better" is not. If you cannot name the defect, you are not ready to constrain—you are about to nudge.
The Discipline
Diagnose is a thinking stage, not a typing stage. Sometimes you diagnose and realize the output is actually fine, which ends the loop. That is a feature: diagnosis doubles as your stopping test, a point developed in Which Numbers Tell You a Refinement Loop Is Actually Healthy.
Stage Three: Constrain
What It Does
Constrain converts a diagnosis into a targeted instruction. The instruction names the defect, supplies the fix or a reference, and pins what must stay fixed.
How to Run It Well
A good constrain turn does three things at once: it states the specific change, it provides a model of the desired result when possible, and it names the invariant—"keep the facts, change only the tone." Each of these prevents a known failure mode.
One Constraint at a Time
On hard problems, apply one constraint per turn so you can tell which instruction worked. Bundling fixes makes the loop opaque, as the spiraling examples in Six Refinement Loops That Turned Mediocre AI Output Into Shippable Work show.
The Advancement Rule
Moving Between Stages
The method's core rule: never constrain without first diagnosing. Every refinement turn must pass through diagnose, even if diagnosis takes three seconds. This single rule eliminates the vague-nudge spiral that kills most loops.
The Loop Shape
A healthy loop looks like Draft once, then Diagnose-Constrain pairs until diagnosis comes up empty. When diagnose returns "nothing meaningfully wrong," you stop. The model has a built-in termination condition rather than relying on willpower.
When to Apply or Adapt the Method
Use the Full Method When
The task is high-stakes, the quality bar is specific, or the output will be reused. Client-facing copy, production code, and analysis that informs a decision all justify the full Draft-Diagnose-Constrain discipline.
Compress It When
The task is low-stakes and disposable—a quick internal summary, a throwaway brainstorm. Here a single draft may be enough, and forcing the full loop wastes effort. The method scales down as well as up.
Scale It Across a Team
Because the stages are named, they are teachable. A team can adopt "always diagnose before you constrain" as a shared norm, which standardizes loop quality regardless of who is at the keyboard.
Worked Walkthrough of the Method
Draft
A consultant needs a one-paragraph summary of a client's quarterly performance for an executive audience. The draft prompt front-loads the context: audience, length, tone, and the three numbers that matter. The first output is a competent but generic paragraph that buries the one figure the executive cares about.
Diagnose
Rather than nudging, the consultant names the defect precisely: the lead sentence opens with a minor metric, and the headline number—a 14-point retention gain—appears only in the last line. The diagnosis is specific enough that the fix is obvious.
Constrain
The constrain turn does the three jobs at once: "Lead with the retention gain. Keep all the numbers exactly as written. Hold the length to one paragraph." A single targeted instruction, naming the change and pinning the invariants. The second output leads with the right figure and meets the bar. The loop ends because the next diagnose pass finds nothing meaningful to fix.
Why the Walkthrough Matters
Notice that no step required clever wording—only specificity. The method's power is that it forces specificity at each stage, which is exactly what undisciplined nudging skips.
How the Method Relates to Other Practices
It Underpins the Checklist
The checklist for running refinement loops is essentially this method turned into a step-by-step verification you can keep open while working. The framework explains why each item exists; the checklist enforces it in the moment.
It Sets Up Measurement
Because the method has a clear stopping condition—an empty diagnosis—it gives you a natural unit to count: passes to acceptance. That metric, central to the numbers that reveal loop health, only makes sense once the loop has defined stages and a defined end.
It Survives Automation
As agentic systems begin running loops autonomously, the method describes what they do internally: generate, critique, constrain, repeat until a stopping condition. Understanding the stages lets you supervise an automated loop and spot when the model gets a stage wrong, which it still does.
Frequently Asked Questions
What makes this a framework rather than just good advice?
The named stages and the advancement rule. Because Draft, Diagnose, and Constrain are distinct steps with a hard rule between them—never constrain without diagnosing—the method is reproducible and teachable, not a matter of individual instinct.
Why is the diagnose stage so emphasized?
Because it is the stage people skip, and skipping it causes the vague-nudge spiral. Forcing yourself to name the defect before instructing both improves the fix and doubles as your stopping test.
How does the loop know when to stop?
When a diagnose pass comes up empty—nothing meaningfully wrong—you stop. The method builds termination into the diagnose stage rather than relying on you to decide you are done.
Can I apply more than one constraint per turn?
On easy problems, yes. On hard ones, apply a single constraint per turn so you can attribute the result. Bundling fixes makes it impossible to tell which instruction actually worked.
When should I not use the full method?
For low-stakes, disposable output. A throwaway internal note rarely justifies a full loop. Reserve the complete Draft-Diagnose-Constrain discipline for work that is high-stakes, reused, or held to a specific bar.
Key Takeaways
- The method splits refinement into three named stages: Draft, Diagnose, and Constrain.
- The core rule—never constrain without first diagnosing—eliminates the vague-nudge spiral.
- Diagnose doubles as the stopping test; an empty diagnosis ends the loop.
- Apply one constraint per turn on hard problems so you can attribute the fix.
- Use the full method for high-stakes or reused work and compress it for disposable output.