01Saying you don't understand an AI's report is the right response

In September 2026 the operator had handed an AI the job of producing explainer videos. When the AI finished, it sent back a report built from numbers. The operator replied with one line: they could not tell what the report was saying.

The AI apologized and tried again in plain words. It had done one thing. The video screen had stayed the same for long stretches and was dull to watch, so it had made the screen change.

The restated report also gave a before and after. Until then, each chapter showed a single slide of text, and that one slide stayed up for about a minute and a half. The narration kept going while the picture sat still. The AI had looked at how 100 popular explainer videos were put together, then rebuilt the videos so that a few images kept moving in step with the narration.

What this episode tests is whether the operator's reply was just a complaint, or a step that anyone working with AI can repeat.

02"I don't understand" shows the AI the gap in its report

A report exists so that its reader can make the next decision. If the reader cannot decide after reading it, the report is not finished, however correct the work behind it may be.

Saying you don't understand is not an admission that you lack knowledge. It tells the AI that this report does not let you decide. Put the two versions side by side and the missing parts become visible.

AspectFirst reportRestated report
What the first sentence holdsNumbers from the workA summary: one thing was done
Why the change was madeNot statedThe screen sat still and was dull
Before and afterCannot be worked outOne slide for 90 seconds → images that move with the narration
What the reader can doNod, or stay silentKeep it, roll it back, or ask for something else

The numbers in the first report were not wrong. What it lacked was an account of what those numbers meant, and a summary of the change that mattered to the reader.

03Approving a report you don't understand leaves you responsible for changes you never saw

A person decides whether to accept the AI's work. If you nod at a report you do not understand and move on, you have approved a change without knowing what it was.

If someone later complains about how the videos look, you cannot explain what was changed. If you want to undo it, you do not know how far back to go. An approval given while pretending to understand is an approval with the decision left out.

Figure 1 Two paths after an unclear report
Pretend tounderstandAsk for arestatementA report full ofnumbersApprove itas isYou don't knowwhat changedand cannot roll itback laterA report full ofnumbersReply: I don'tunderstandOne-sentencerestatementwhat changed andwhyDecide to keepor roll backPretend to understandAsk for a restatementA report full ofnumbersApprove itas isYou don't knowwhat changedand cannot roll it backlaterA report full ofnumbersReply: I don'tunderstandOne-sentencerestatementwhat changed and whyDecide to keep orroll back
On the top path you approve without knowing the content. On the bottom path one reply gives you what you need to decide.

Replying that you don't understand costs one line. That line gets you the material for the decision. The effort saved by not asking is small next to what can be lost later.

04Ask for four parts, with the change stated in the first sentence

A plain restatement, as used here, is a report that someone without the specialist vocabulary can read and then decide whether to keep or roll back the work. The restatement the operator received had four parts.

1

What changed

One sentence, first

State the whole first, as in: one thing was done, the screen now changes.

2

What the problem was

The reason for the change

A problem the reader cares about: the screen sat still and was dull.

3

Before and after

In a form you can compare

One slide held for 90 seconds, then the rebuilt version.

4

Numbers

Last, as evidence

How many videos were studied, how often the text changes. Each with its meaning.

A restatement does not throw the numbers away. It moves them from in front of the summary to behind it. They stay as the evidence you use to check whether the summary is true.

This episode covers the case where you cannot follow the point of the whole report. When only one word in a report is unclear, it is enough to ask what that word means on the spot. Episode 8 of this series covers that.

05In material review and medical work, ask for a report you can decide on

When a pharmaceutical company has an AI review a draft promotional material, the report can come back as a count of edited passages, a match rate against internal rules, and a list of changed phrases. The numbers may be complete. What the reviewer needs to know is what the material now claims.

The same holds when a medical affairs specialist has an AI summarize a paper's analysis. Before the rows of figures, have the AI state in one sentence what the result supports and what it does not. Then you can decide whether the summary is fit for an internal decision.

The pharmaceutical field has already turned this idea into a rule. The EU Clinical Trials Regulation (Regulation (EU) No 536/2014) requires the sponsor to submit, along with the summary of results, a summary that a layperson can understand. It treats the numerical report and the plain-language report as two separate things, both required.

Government writing guidance says the same. In the United States, the Plain Writing Act of 2010 requires public-facing content to be written for its specific audience, and Digital.gov publishes guides on how to do it. In Japan, the Council for Cultural Affairs issued guidance in 2022 asking writers to replace specialist terms or explain them for the reader. Asking the same of an AI's report is not a special demand.

06AI reports fill with numbers because they are written from inside the work

During a task, an AI handles one value after another: counts, durations, settings. When the task ends, the report tends to be written from that record of values. It speaks in the language of the work, and the change on the reader's side comes last or not at all.

People show the same tendency. Someone who knows a subject is poor at judging how a person who does not know it will take in information. The economist Colin Camerer and colleagues showed this in a 1989 paper, and the effect later came to be called the curse of knowledge. Whoever did the work tends to assume the reader shares their starting point.

The design of the AI plays a part too. Anthropic's official guidance for Claude says recent models write more concisely and may skip summaries after work that involves tools. The same guidance says that if you want a summary, you can instruct the model to give a short account of what it did once the task is done.

Figure 2 Three reasons AI reports fill up with numbers
The AI finishes andreportsIt reports the values ithandledcounts, seconds, settingsIt assumes the readerknows the samethe insider's blind spotIt is built to be briefand skips summariesit will write one if askedThe AI finishes and reportsIt reports the values it handledcounts, seconds, settingsIt assumes the reader knows the samethe insider's blind spotIt is built to be brief and skipssummariesit will write one if asked
All three happen on the AI's side and have nothing to do with the reader's knowledge, which is why one reply from the reader fixes them.

Each of these causes sits on the AI's side. None comes from a gap in the reader's knowledge. That is why one reply from the reader can fix it.

07From tomorrow, reply that you don't understand and have the AI lead with the change

There are four steps.

  1. Say you don't understand. Add where you lost the thread, and the restatement comes faster. If it is the whole report, say so.
  2. Have it state what changed in the first sentence. Ask for "what you did, in one sentence, without specialist terms."
  3. Have it give before, after, and the problem. Ask what was wrong and what is now different.
  4. Have it put the numbers last, with their meaning. Each number should come with a few words on what it is there to check.

If one restatement still leaves you lost, reply again. Once you have a report you can follow, put that format at the top of your future requests. Then you will not have to ask every time.

Figure 3 Four steps to a usable restatement
Say you don'tunderstandand from whereWhatchanged,…Before,after, and…Numberslast, as…Stillunclear?Say you don't understandand from whereWhat changed, first sentenceBefore, after, and the problemNumbers last, as evidenceStill unclear?
If one restatement is not enough, reply again. Once it is clear, add that format to your standing instructions.
Key Points ── 3 to take away
  1. When you cannot follow an AI's report, the cause is a missing summary and missing meaning in the report, not a lack of skill on your side. You can say you don't understand.
  2. Approving a report you don't understand leaves you responsible for changes you never saw. One line of reply gets you what you need to decide.
  3. Ask for the restatement in this order: what changed, what the problem was, before and after, then the numbers. Keep the numbers, but as evidence at the end.
Closing

When an AI's report leaves you lost, you do not have to nod along. A report is finished only when its reader can decide. Reply that you don't understand, have the AI lead with what changed, and the choice to keep or roll back its work stays with you.

Sources & references
  1. Anthropic. Prompting best practices (Communication style and verbosity). Claude Developer Platform Docs. https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/claude-prompting-best-practices
  2. Council for Cultural Affairs (Japan). Kōyōbun sakusei no kangaekata [Approach to writing official documents]. Agency for Cultural Affairs, 2022. https://www.bunka.go.jp/seisaku/bunkashingikai/kokugo/hokoku/pdf/93651301_01.pdf
  3. European Parliament and Council. Regulation (EU) No 536/2014 on clinical trials on medicinal products for human use. Official Journal of the European Union, 2014. https://eur-lex.europa.eu/eli/reg/2014/536/oj
  4. U.S. General Services Administration. Plain Language Guide Series. Digital.gov. https://digital.gov/guides/plain-language
  5. Wikipedia. Curse of knowledge (includes Camerer, Loewenstein and Weber, Journal of Political Economy, 1989). https://en.wikipedia.org/wiki/Curse_of_knowledge
What this episode is based on The operator's own record of requests to and decisions with Claude since February 2026 (the operator has used generative AI since March 2023), anonymised and generalised into a pattern. No messages are quoted. The sources listed are public material used to check the background of the pattern.