The limit of a perfect instruction
There is a comforting belief that if you just word the request well enough, you will get the right answer first time. Be specific. Give context. State the format. All good advice, and it does lift the quality of a single answer.
It also runs out. Wording is something you decide before anything has happened. There are whole categories of problem where the useful information does not exist yet at the moment you press send, and no amount of care with your sentences can conjure it into being.
โ Weak prompt
Prompt
Write a competitor comparison of the three main tools in our category. Be thorough, accurate, use a table, cite sources, cover pricing, features and support.
Output
A confident, tidy, well-formatted table covering three tools, with pricing tiers and feature ticks laid out neatly across all three columns.
The instruction is excellent and the output looks excellent. But nothing checked whether those prices are current, and one of them changed in April. A single pass had no opportunity to find that out.
โ Good prompt
Prompt
Same task, but work in passes. First list what you need to look up. Then look each one up. Then check anything that looks out of date against a second source. Then write the table, and mark anything you could not confirm.
Output
A table where two prices are confirmed against the vendor's own page, one is flagged as changed recently, and a support claim is marked as unverified.
Same underlying tool, same knowledge. The difference is that something got to look at the intermediate results and react to them.
What the second pass actually buys you
Three things, and they are worth separating because they fail differently.
It catches its own mistakes. Producing an answer and judging an answer are different jobs, and the second is easier. A tool that has just written a summary can be asked to compare that summary against the source and list what it missed. It will often find real gaps, because reviewing is a smaller task than writing.
It breaks a big job into pieces that fit. Ask for a fifteen page report in one go and you get something thin and even, because attention was spread across everything at once. Ask for an outline, then each section against that outline, then a pass to make the sections agree with each other, and each individual step is small enough to do properly.
It adapts when something unexpected turns up. This is the one wording cannot fake. The file is empty. The page has moved. The data contradicts the assumption in your brief. A single pass either ignores that or invents its way past it. A loop notices and changes course.
In plain English
- Self-check:
- A pass where the AI reviews its own previous output against the brief instead of producing new work.
- Decomposition:
- Splitting one large job into smaller pieces that get handled one at a time.
- Feedback:
- Any information coming back from an action: an error, a result, a file that was not what anyone expected.
- Verification:
- Checking a claim against something outside the AI, such as a document or a live page.
BRIEF Write a 15-page onboarding guide for new hires.
ONE PASS
Produces 15 pages in one continuous run.
Every section gets roughly equal attention, whether it needs it or not.
Section 4 contradicts section 11, because nothing ever compared them.
A policy detail is stated confidently and is two years out of date.
BRIEF Write a 15-page onboarding guide for new hires.
PASS 1 Draft the outline. Ask which sections need source documents.
PASS 2 Write sections 1 to 5 against the outline.
PASS 3 Write sections 6 to 15.
PASS 4 Read all sections together. Find contradictions.
Found: section 4 and section 11 disagree on the probation period.
PASS 5 Check the probation period against the HR policy document.
Found: both were wrong. Fix both.
PASS 6 Final read against the original brief.
Same tool. Same knowledge. The difference is that
something looked at the work before you did.
Why this is not just being fussy
The instinct is to see extra passes as perfectionism. It is closer to the opposite. Each pass is a cheap chance to find out that you are wrong before the wrongness has been built on.
Compare it to how a decent piece of work moves through a team. Someone drafts. Someone else reads it. A fact gets checked. A section gets rewritten because the reader misread it, which is a signal that the writing was unclear. Nobody thinks of this as fussiness. It is just how work that has to be right gets made.
The loop is that process compressed, run by one tool, at a speed that makes it practical to do many more times than a team ever would.
This is also why you sometimes see an AI tool visibly change its mind partway through a job. That is not confusion. It is the check step doing what it exists to do. A tool that never changes course is not more reliable, it just has no mechanism for noticing when it is off.
Do this in two stages and show me both.
STAGE 1: produce your best attempt at [the task].
STAGE 2: read your own stage 1 output back against my brief.
List anything that is missing, unsupported, or that you are not
confident about. Then give me a corrected version.
Do not skip stage 2 even if stage 1 looks fine.
You can run this loop by hand in any chat tool. Ask for the answer, then ask what is weakest about it, then ask for the fix. Three messages, and it usually beats one very carefully worded message.
Checkpoint
Iteration buys three things wording cannot: the chance to catch its own mistakes, a way to break big jobs into pieces small enough to do properly, and the ability to adapt when reality does not match the brief.
Quick Quiz
๐ Quiz
Question 1 of 4Why can better wording alone not replace a loop?