01The operator stated the same order twice in one day, in different words

The material is two requests from September 9, 2026. In the first, the operator asked the AI to move on to building the English site, one step after the other, once the Japanese version was done.

The second was sent as a request to be picked up later. It opened by saying this hardly needed saying, repeated that English came next, and asked that English begin only after the Japanese go-live had fully finished.

The limits of this material matter. Only the requests survive here. These two records cannot tell us whether the AI ever broke the order, or why the operator wrote the second request. What they do show is one order written twice in a day, with a more precise condition the second time.

That is enough to extract a pattern anyone can use. When you feel the urge to restate an order, first look at what the last version left out.

02The second request added a start condition, not a new order

Split the two requests into parts and the order itself has not moved. Japanese comes first and English second. What changed is how the moment to start English is described.

AspectFirst requestSecond request
Earlier taskJapanese versionJapanese version
Finish conditionOnce it is done (which state is not stated)After go-live has fully finished
Next taskBuilding the English siteThe English version
Run side by side?No, one after the otherNo, one after the other

"Done" has a wide range. It could mean the draft is written, the page is assembled, or the page is published. "Go-live has fully finished" leaves far less room for different readings.

Figure 1 Readings of "done" and the point where English may start
DraftwrittenPageassembledPagepublishedGo-liveconfirmed…Start theEnglish…Draft writtenPage assembledPage publishedGo-live confirmedfinished?Start the English version
"Once it is done" can mean any of the first three. The second request fixed the starting point after go-live had finished.

03A vague start condition lets corrections miss the next task

Order rules matter when the later task builds on the result of the earlier one. Correct the Japanese just before publication, and an English version already under way will not contain that correction. The same passage gets fixed twice, and anything missed remains as a mismatch between the two versions.

Aligning the versions shows up in the record too. On the same September 9, the operator separately asked for a job that would find and resolve differences between the Japanese and English pages. The record does not link that job to the order rule. It does show that keeping two versions aligned had become a separate task of its own.

The other cost is the user's attention. A rule that lives only in your head has to be remembered and typed each time, and on the day you forget, it might as well not exist. An order you consider too obvious to mention still reaches the AI only as far as you wrote it down.

04An order rule is one sentence with four parts

In this episode, an order rule means a promise that holds repeatedly, of the form "this task comes after that one", written as one sentence with these four parts.

1

Earlier task

What comes first

Name the work the next step builds on, such as publishing the Japanese version.

2

Finish condition

What state counts as finished

Describe a state someone can check and answer yes or no.

3

Next task

What starts then

Name the work that begins once the condition is met.

4

Not alongside

What waits until then

State plainly what must stay untouched until the condition is met.

The second part carries the most weight. Anthropic's Claude Code documentation recommends instructions concrete enough to verify, and offers "Use 2-space indentation" in place of "Format code properly". For an order, that means "after checking the published page" in place of "when finished".

The idea has a boundary. An instruction that sets the sequence for one piece of work, once, is not a rule in this sense. Forcing an order on independent tasks slows down work that could run in parallel. Make a rule only when the later task builds on the earlier result and the same order keeps coming back.

05In material review, it applies when derivatives follow a settled source

Pharmaceutical material production contains orders of the same shape. The table below applies the pattern as an illustration. Actual conditions should follow each company's own procedures.

SituationVague wordingWording that works the first time
Source material and English versionTranslate it once the Japanese is readyStart the English draft after approval of the Japanese material is recorded in internal review. Do not translate before then
Brochure and briefing slidesDo the slides when the brochure settlesBuild the slides from the final brochure once it is confirmed, and state the version number used
Package insert revision and materialsFix the materials when the revision arrivesAfter receiving the revised insert, list the affected materials. Do not edit any text until the person responsible has checked the list

In each example, the AI's part ends at a draft or a list. Whether something is approved is for the reviewer to decide, not the AI. Writing a human decision into the finish condition lets you instruct the AI not to move past it.

06An order given in conversation may not last until the next task

Product design helps explain why orders get restated. Take Claude Code as the example. Its documentation says each session begins with a fresh context, and that an instruction given only in conversation can disappear after a long conversation is compacted.

Writing it in a settings file does not make it absolute either. The same documentation says the AI treats that file as context rather than enforced configuration, and that strict compliance is not guaranteed, especially for vague or conflicting instructions.

Figure 2 What can happen to an order given only in conversation
Order given only inconversationNot carried to a newsessionfresh context each timeCan vanish aftercompactionin long conversationsReadings divergewhen the condition is vagueOrder given only in conversationNot carried to a new sessionfresh context each timeCan vanish after compactionin long conversationsReadings divergewhen the condition is vague
Based on Claude Code's documentation. Other products work differently, so check the documents for your own AI.

So choose the location by what a broken order would cost. Claude Code's documentation advises that anything that must happen every time should be enforced by the tooling, not left as an instruction the AI weighs. Software automation offers the same idea. In GitHub Actions, you can declare that a job runs only after another job succeeds. If the earlier job fails, the later one is skipped.

Different products work differently. Check your own AI's documentation to learn what carries over to the next conversation.

07From tomorrow: turn an order you stated twice into a sentence, then choose its home and its stop

There are four steps.

  1. Find the order you restated. Claude Code's documentation also lists anything you have had to correct twice as a candidate for the always-loaded file. Treat the second time as the signal to look for what the first version lacked.
  2. Write it as a four-part sentence. Use this form: "After X reaches (a state you can check), start Y. Leave Y untouched until then." Add the reason in a few words. Anthropic's documentation explains that giving the motivation behind an instruction helps produce responses better aimed at your goal.
  3. Keep it where it loads every time. If your AI has project instructions or an always-read settings field, put it there. If it does not, make it part of your standard request text.
  4. Add a stop to orders that are hard to undo. If publication or sending outside the company is involved, place a human approval or an automated check before the next task.
Figure 3 From restated order to a rule in its place
Same orderstated twiceWrite onefour-part…Keep it whereit loads…Hard toundo if…Add humanapproval or a…Same order stated twiceWrite one four-part sentenceKeep it where it loads every timeHard to undo ifbroken?Add human approval or a check
If a broken order can be undone, the sentence is enough. Orders involving publication or outside sending also need a stop.

Then test it once. In a new conversation, ask the AI what it may start next, and have it state the condition in its own words. If the answer drifts, fix the wording at that point. A single sentence does not guarantee compliance, so the check before publication stays with a person.

Key Points ── 3 to take away
  1. When you restate an order, look at the start condition before repeating yourself. Replace "when done" with a state that can be checked.
  2. An order rule is one sentence with four parts: earlier task, finish condition, next task, and what must not run alongside. Write one only for orders that have a real dependency and keep coming back.
  3. Instructions given in conversation may not last. Keep the rule where it loads every time, and add a human approval or an automated check to orders that are hard to undo.
Closing

Repeating an order to an AI is not only a question of what the AI retains. It also signals a finish condition that is still in your head and not yet on the page. Write the condition down and give it a home, and the next request can start without retyping the order. Confirming that the order was kept remains the user's job.

Sources & references
  1. Anthropic. How Claude remembers your project. Claude Code Docs. https://code.claude.com/docs/en/memory
  2. Anthropic. Best practices for Claude Code. Claude Code Docs. https://code.claude.com/docs/en/best-practices
  3. Anthropic. Prompting best practices. Claude Developer Platform Docs. https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/claude-prompting-best-practices
  4. GitHub. Using jobs in a workflow. GitHub Docs. https://docs.github.com/en/actions/how-tos/write-workflows/choose-what-workflows-do/use-jobs
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.