Nothing about this is free
A loop feels free because you are not doing the work. The tool is busy, you are not, and the natural instinct is to let it run. But every pass spends three real things.
Money. Each pass is more work done on your behalf, and whether you pay by subscription, by usage, or through a limit on how much you can do per day, it comes out of something.
Time. Not the tool's time, yours. The job is not finished until it is finished, and while it runs you are either waiting or context-switching, which is its own tax.
Attention. This is the one people miss. The longer the run, the more work there is for you to review at the end. A loop that took many passes has produced more decisions to check, and the reviewing does not get delegated.
In plain English
- Diminishing returns:
- Each extra pass improves the work less than the one before it, while costing the same.
- Fiddling:
- Changes that alter the work without improving it, often reversing an earlier change.
- Cost per pass:
- What one trip round the loop costs you in money, time and review effort combined.
- Good enough:
- The point where the work meets the brief and further passes are buying you nothing.
The shape of improvement
Passes do not improve work evenly. The early ones do nearly all the lifting.
Pass one produces something usable and shapeless. Pass two catches the actual errors, and this is usually the single most valuable pass in the entire run. Pass three tightens the structure and fixes what pass two disturbed. Pass four is a decent polish.
Then something changes. Passes five onwards are still doing work. The document keeps changing. But the changes stop being corrections and start being preferences: a sentence reworded, then reworded back, a section moved and returned. The cost per pass has not dropped at all. The value has.
PASS 1 shapeless but usable draft big improvement
PASS 2 real errors found and fixed biggest improvement
PASS 3 structure tightened solid improvement
PASS 4 polish modest improvement
PASS 5 wording changes hard to tell
PASS 6 some wording changed back none
PASS 7 heading renamed none
PASS 8 heading renamed again none
The cost of pass 8 is the same as the cost of pass 2.
How to spot the switch from fixing to fiddling
You do not have to watch every step. There are four tells, and any one of them is enough.
Changes start reversing. Something was changed in one pass and changed back in another. That is the clearest signal available, because a loop undoing its own work has stopped learning anything new.
The reasons get vaguer. Early passes report specific findings: this figure does not match the source, this section is missing. Late passes report improved clarity and refined the tone. Vague reasons usually mean nothing concrete was found.
The diff shrinks. If you can look at what changed between two versions and struggle to see a difference that matters, you have your answer.
You stop being able to describe the gain. Ask yourself what the last three passes bought. If you cannot say in one sentence, they bought nothing.
The reversal tell is worth trusting above the others. A loop that changes something and then changes it back is not converging on a better answer, it is oscillating between two acceptable ones. Neither is going to win, and every pass costs the same as the ones that were genuinely useful.
Deciding before you start
The cheapest way to control loop cost is to decide how much the job is worth before the loop begins, not while you are watching it.
Some jobs deserve very few passes. A first draft you are going to rewrite anyway needs one. An internal note that three people will read needs two. Some deserve many: anything a customer will see, anything with numbers in it, anything you will be quoted on later.
The question is not how good could this be. It is how much does being wrong here cost me. Answer that first, and the number of passes falls out of it.
LOW STAKES internal note, rough draft, personal reminder
one or two passes, then stop
MEDIUM STAKES team document, client email, meeting summary
a few passes, one of which must be a fact check
HIGH STAKES anything customer-facing, anything with figures,
anything you will be held to
more passes, plus a human read at the end
that is not optional
The extra passes on a high stakes job are cheap.
The extra passes on a low stakes job are the whole cost.
Work on this in at most [4] passes.
After each pass, tell me in one line what that pass changed
and why it was needed.
If a pass would only change wording or reverse an earlier change,
stop instead and tell me you have reached good enough.
Do not use a pass to improve something that already meets the brief.
Asking for a one line reason per pass costs almost nothing and makes fiddling visible immediately. When the reasons start reading like refined the phrasing, you have your stopping point without having to watch anything.
The quiet cost of a long run
There is one more cost, and it does not appear on any bill. The longer a loop runs unwatched, the harder its output is to review, and the more tempted you are to accept it because reviewing it properly looks like more work than it was worth.
That temptation is how drifted work gets shipped. A short run is not just cheaper. It is easier to check, which makes it likelier that you actually check it.
Checkpoint
Every pass costs the same, but the early passes deliver most of the value. When changes start reversing and reasons get vague, the loop has moved from fixing to fiddling. Decide how many passes the job deserves before you start.
Quick Quiz
๐ Quiz
Question 1 of 4Which pass typically delivers the most value in a run?