Knowing that refinement works is different from having a routine you can run under pressure. A routine tells you what to do first, what to do when a draft comes back wrong in a specific way, who is responsible for each step, and in what order it all happens. Without that, even people who understand refinement improvise inconsistently and get inconsistent results.
This piece lays out an operating routine for iterative refinement as a set of named plays. Each play has a trigger that tells you when to run it, an action, and a clear owner. The plays are sequenced so that you tackle the expensive decisions before the cheap ones and never waste a pass on the wrong layer of the work.
Treat this as something you run, not just read. The value is in following the sequence the next time you have a real draft in front of you, rather than reinventing your approach every time.
The Setup Plays: Before the First Draft
The most important plays happen before any text exists. Skipping them is why most loops sprawl.
Play one: write the acceptance criteria
Trigger: you are about to start any output worth refining. Action: write down what done looks like as a short list of named, checkable criteria. Owner: whoever is accountable for the final output. This is the single highest-leverage play in the routine, because it gives every later step something fixed to aim at. The framework piece offers criteria structures by work type.
Play two: lock the structure
Trigger: criteria are set, before drafting content. Action: have the model propose a skeleton, an outline, an interface, a plan shape, and approve it before any detail is generated. Owner: the author. Locking structure early prevents the most expensive kind of late pass, the restructure, which often undoes hours of detail work.
The Generation Plays: First Draft
With setup done, generation becomes cheap and low-stakes, because the loop already knows where it is going.
Play three: generate against the locked structure
Trigger: structure approved. Action: produce the first full draft inside the agreed skeleton, with the acceptance criteria visible in the prompt. Owner: the author. The draft does not need to be good; it needs to be complete enough to critique. The getting started guide covers a clean first generation.
The Critique Plays: Finding the Real Gaps
Critique is where the routine earns its results. These plays run every pass.
Play four: critique against the rubric, in a fresh context
Trigger: a draft exists. Action: evaluate the draft against each acceptance criterion, marking pass or fail with a one-line reason, in a context separate from the one that generated it. Owner: a human, optionally assisted by the model. Running critique fresh keeps it honest. The advanced piece details why separating critic from author matters.
Play five: rank the failures by impact
Trigger: critique findings are in hand. Action: order the failed criteria by how much they affect the output's worth, and select the top one or two to address this pass. Owner: the author. Fixing everything at once invites new problems; fixing the most important thing first keeps each pass focused.
The Revision Plays: Closing the Gap
Revision converts critique into a better draft without introducing new faults.
Play six: revise only the selected failures
Trigger: top failures selected. Action: instruct the model to correct specifically those items, leaving everything passing untouched. Owner: the author. Surgical revision beats wholesale rewriting, which tends to undo good work while fixing the bad. The best practices piece covers writing directive corrections.
Play seven: re-run critique on the revised draft
Trigger: a revision exists. Action: re-evaluate against the full rubric, confirming the fix landed and nothing else regressed. Owner: the critic. This catches the common failure where fixing one thing quietly breaks another.
The Closing Plays: Knowing When to Stop
The routine ends decisively rather than trailing off into endless polishing.
Play eight: stop when the rubric passes
Trigger: every acceptance criterion passes. Action: ship. Owner: whoever is accountable for the output. Resist further tweaking; once the criteria are met, additional passes are polishing past value. The risks piece treats over-refinement as the trap it is.
Play nine: capture what you learned
Trigger: output shipped. Action: note any criterion that was missing or any recurring fault, and fold it into the rubric for next time. Owner: the author or standards owner. This turns each loop into a small improvement to the routine itself.
Running the Plays Under Different Conditions
The routine adapts to circumstance. Knowing how to flex it is part of running it well.
The fast variant for routine work
When the output is low-stakes and familiar, collapse the routine without abandoning its spine. Use a saved rubric and a saved structure, generate once, run a single focused critique-and-revise pass, and ship. The setup plays still happen; they just draw on prior work rather than fresh thought. This keeps quality consistent on volume work without ceremony.
The deep variant for high-stakes work
When the output carries real consequence, expand the routine. Add a second independent critique from a different person or a fresh model context, widen the acceptance criteria to cover risk and edge cases, and keep a record of major changes per pass for accountability. The deep variant trades speed for defensibility, which is the right trade when the cost of being wrong is high.
Adapting when the first draft is off-strategy
The routine assumes the first draft is roughly right in approach and imperfect in execution. When a draft is fundamentally off-strategy, do not run the revision plays on it; that only refines a wrong thing. Return to the setup plays, re-examine the criteria and structure, and regenerate. Recognizing this early is itself a play, triggered when two passes have not improved the core.
Frequently Asked Questions
Which play matters most if I only adopt one?
Writing the acceptance criteria before drafting. Every other play depends on having a fixed target to aim at and check against. Without it, critique has no standard, revision has no direction, and stopping has no trigger. It is the play that converts open-ended polishing into a bounded, convergent routine.
How do I sequence the plays under deadline pressure?
Keep the setup plays even when rushed, because they save more time than they cost, then compress the loop to fewer passes rather than skipping critique. A fast run is criteria, locked structure, one draft, one focused critique-and-revise, ship. Cutting setup to save minutes is what makes loops sprawl and blow deadlines.
Who should own the critique play?
A human, ideally not the same person who is emotionally invested in the draft, with optional model assistance. Ownership matters because critique is where the standard is enforced. On a team, this often means a peer or reviewer applies the rubric, which keeps the assessment honest and the standard consistent.
Can this routine be run by one person?
Yes. A solo practitioner plays every role but should still separate the critique context from the generation context to keep the assessment honest. The discipline of switching hats, author, then fresh critic, then author again, preserves most of the benefit of having a separate reviewer.
How is this routine different from just iterating?
Iterating is improvising; the routine is sequenced and triggered. It puts the expensive decisions, criteria and structure, before the cheap ones, assigns owners so nothing falls through, and ends with an explicit stop. The structure is what makes results consistent instead of dependent on the operator's mood that day.
How do I keep the routine from feeling bureaucratic?
Keep the artifacts lightweight: a short criteria list, not a document; a quick pass-fail critique, not a report. The routine should add minutes, not meetings. If it starts feeling heavy, you are over-formalizing; strip it back to the named plays and their triggers and let the rest be judgment.
Key Takeaways
- The highest-leverage plays happen before drafting: write acceptance criteria and lock structure first.
- Run critique in a fresh context, rank failures by impact, and fix only the top one or two per pass.
- Revise surgically rather than rewriting wholesale, then re-run critique to catch regressions.
- End the routine decisively when the rubric passes, and capture lessons back into the rubric for next time.
- The routine works solo or on a team; its value is sequencing and clear ownership, not added bureaucracy.