Iterative refinement is widely recommended and widely misunderstood. People hear "iterate" and form a picture that is partly right and partly a set of beliefs that actively undermine results. They assume more passes are always better, that any iteration is refinement, that the model can reliably grade itself, or that a clever enough first prompt makes the loop unnecessary. Each of these sounds reasonable. Each is wrong in a way that costs time and quality.
The trouble with these myths is that they are close enough to true to survive casual scrutiny. More passes often do help, until they do not. The model often can spot its own flaws, until it cannot. Believing the comfortable half of each truth leads people to run loops that waste effort or, worse, degrade the very work they are trying to improve.
This piece takes the most common misconceptions one at a time and replaces each with the accurate picture, grounded in how loops actually behave rather than how they are imagined to.
Myth: More Passes Always Mean Better Output
The most pervasive belief is that quality scales with iteration, so if two passes are good, five must be better.
The reality of diminishing and negative returns
Quality improves with passes only up to convergence, after which additional passes change wording without changing worth, and beyond that can introduce new problems while fixing trivial ones. The curve is not monotonic. A loop that runs too long can produce a fifth draft that is genuinely worse than the second.
The accurate picture is that passes have a sweet spot, usually two to four for most tasks, and the skill is knowing where it is. The metrics piece covers how to recognize convergence rather than chasing it.
Myth: Any Iteration Counts as Refinement
A second misconception treats refinement and "trying again" as the same thing.
Reactive retrying versus directed refinement
Generating a new draft because you did not like the last one is not refinement; it is rerolling. True refinement holds a fixed target, identifies the specific gap between the current draft and that target, and corrects that gap. Without a stable specification, you are not converging, you are sampling, and sampling rarely produces excellence.
The accurate picture is that refinement requires a defined standard and a specific critique. Anything less is rolling dice and stopping when you like the result. The framework piece draws this distinction sharply.
Myth: The Model Can Reliably Judge Its Own Work
Because models can critique outputs, people assume they can be trusted to grade themselves and decide when the work is done.
Where self-assessment breaks
A model in the same context that produced a draft tends to defend that draft. Even with a fresh context, its judgment is only as good as the criteria you give it; with vague criteria, its self-assessment is confident and unreliable. It will declare done before done, or flag problems that are not problems.
The accurate picture is that model critique is a useful assistant, not a trustworthy judge. The human owns the standard and the final call, especially on anything that matters. This connects directly to the risk of outsourcing your judgment entirely.
Myth: A Perfect First Prompt Makes Loops Unnecessary
Some people believe loops are a crutch and that a sufficiently engineered initial prompt eliminates the need to iterate.
Why one-shot fails on hard problems
For multidimensional outputs with competing constraints, you usually cannot fully specify what you want in advance, because you discover requirements by seeing drafts. The first draft reveals what the specification missed. No prompt, however careful, surfaces those gaps before generation.
The accurate picture is that loops and good first prompts are complements, not substitutes. A strong first prompt shortens the loop; it does not remove it. The best practices piece treats prompt quality and loop discipline as two halves of the same skill.
Myth: Refinement Is Only for Long-Form Writing
A final misconception scopes refinement too narrowly, treating it as an editing trick for essays and articles.
The broad applicability
The generate-critique-revise loop applies anywhere model output has multiple quality dimensions: code, analysis, plans, designs, structured data, and decisions. Restricting it to prose means missing most of its value. The accurate picture is that refinement is a general method for converging on quality, not a writing-specific technique.
Myth: Refinement Is the Slow, Inefficient Option
A final cluster of misconceptions treats iteration as a luxury you indulge when you have time to spare.
The false trade-off with speed
People imagine a choice between fast one-shot output and slow refined output, and under deadline they pick fast. But a disciplined loop with a fixed target and a pass budget often reaches a usable result faster than repeatedly tweaking a single prompt and hoping. Undirected fiddling is the slow path; structured refinement is frequently the quicker route to something you can actually ship.
Where the time actually goes
The accurate picture is that refinement front-loads a small amount of time, writing acceptance criteria, locking structure, and saves a larger amount later by preventing the expensive rework that comes from shipping an under-baked first draft and fixing it after the fact. The visible cost of iterating is more than offset by the invisible cost it avoids. The best practices piece treats this front-loading as central rather than optional.
The myth that bounds matter only for experts
Beginners often believe pass budgets and stopping rules are advanced refinements they can skip. The opposite is true: novices benefit most from explicit bounds, because they lack the instinct that tells an expert when to stop. Structure is not the reward for skill; it is the scaffold that produces skill.
Frequently Asked Questions
Is it true that more refinement passes always improve the result?
No. Quality improves only up to convergence, typically within two to four passes for most tasks. Beyond that, passes change wording without changing worth and can introduce new flaws while fixing trivial ones. The output curve is not monotonic, so blindly adding passes can make the work worse, not better.
Can I just trust the model to tell me when the work is done?
Not reliably. A model judging its own output, especially in the context that produced it, tends to defend that output, and its assessment is only as good as the criteria you supply. Treat model critique as a useful assistant, but keep the standard and the final decision in human hands.
If I write a good enough first prompt, can I skip the loop?
Usually not for hard tasks. You discover real requirements by seeing drafts, so the first output reveals gaps no prompt could anticipate. A strong first prompt shortens the loop but does not eliminate it. Loops and good prompts are complements; treating them as substitutes leaves quality on the table.
Is rerolling the same as refining?
No. Generating a new draft because you disliked the last is sampling, not refinement. True refinement holds a fixed target, identifies the specific gap between the draft and that target, and corrects it. Without a stable standard and a specific critique, you are rolling dice rather than converging on quality.
Does iterative refinement only apply to writing?
No. The generate-critique-revise loop works anywhere output has multiple quality dimensions, including code, analysis, plans, designs, and structured data. Scoping it to prose is itself a misconception that hides most of its value. It is a general method for converging on quality, not a writing-specific editing trick.
Why do these myths persist if they are wrong?
Because each is close enough to true to survive casual use. More passes often do help, models often do spot real flaws, and good prompts often do reduce iteration. People generalize the comfortable half of each truth into a rule, and the rule fails only in the cases that matter most, which are easy to rationalize away.
Key Takeaways
- Quality scales with passes only to convergence, then flattens and can reverse; the sweet spot is usually two to four passes.
- Rerolling is not refining. Refinement requires a fixed target and a specific critique, not just trying again.
- A model is a useful critic but an unreliable judge of its own work; humans must own the standard and the final call.
- Good first prompts and refinement loops are complements; a strong prompt shortens the loop but rarely removes it.
- Refinement is a general convergence method that applies to code, analysis, and design, not just to prose.