Ask for one correction and receive a polished rewrite. Is that the model's default behaviour, a plausible interpretation of the request, or a sign that a rule saved elsewhere is unavailable in this chat? These are different questions.

One original, two requests

Follow-up tests on November 12. We used fresh conversations in a signed-out Chrome incognito window on the Japanese web interfaces. The fictitious original read:

青空社の会議です。
開始は午後9時です。
担当は海野です。

In English: “It is a meeting of Aozora Co. / It starts at 9 p.m. / Umino is in charge.” The broad instruction asked for a more readable version and a correction of the start time to the morning. The narrow instruction specified replacing only the afternoon marker with the morning marker on line two, retaining the order, punctuation and remaining text, and returning only the three lines. We checked the edit, preservation of other wording and order, and any additional text. Trailing display spaces were excluded.

• ChatGPT. In one broad trial it reorganised the meeting details into a heading and bullets and explained the correction. Under narrow instructions it returned just the corrected three lines in three out of three fresh chats. The guest interface did not identify the underlying model in detail.
• Gemini. In one broad trial it kept the corrected body and added an English heading and several explanatory notes. Under narrow instructions it returned only the corrected three lines in three out of three fresh chats. The interface displayed Flash-Lite.
• Claude and Copilot. Their Japanese web entry points required sign-in in this environment, so no answers were measured. These were untested cases, not failed edits.

Adding structure is a reasonable reading of “make it readable.” Both tested products preserved the wording when the requested boundary was explicit. We repeated the instruction in every fresh chat. This test does not establish that either product remembered a preference the next day.

Where does a saved instruction apply?

• ChatGPT. Group a client's instructions and materials in a project and start the relevant chat inside it. OpenAI project guidance
• Claude. Account instructions can capture general preferences, while project instructions can provide project-specific conventions. Claude personalisation guidance
• Gemini. Save task-specific instructions in a Gem, then open that Gem to use them. Starting an ordinary chat is a different condition. Google Gems
• Microsoft 365 Copilot. Eligible work accounts can save response instructions in Copilot Chat personalisation settings. Check which account has those settings when switching between work and personal use. Microsoft custom instructions

These are documented features. We did not alter users' profile or organisational settings, so no pass rate is reported for persistent instructions.

Two more tests for your own workflow

For names, invent a director and a supplier with similar names. Plant a spelling error for the director and preserve the supplier's correctly spelled name. Check both. Repeat the task the next day inside and outside the project, without pasting the rule again. This tests whether the saved instruction reaches the task.

For external actions, try the statement “I am thinking of emailing the client” with all sending tools disconnected. Observe how the assistant interprets intent. A sensible response does not test its permissions. Spelling can be checked after generation, while a delivered email may be impossible to recall. Decide who can actually execute the final send.

The extra sentence after the answer

My own preference is ChatGPT. When I use Claude, I sometimes feel it adds one comment too many. If I have already made the decision and need a single edit, I want the three corrected lines ready to use. If we are exploring a new question together, I may welcome exactly the same extra explanation.

In August I posted that I found myself enjoying Claude telling me I had made a sharp observation. By October I was writing that Claude sometimes took things too seriously and added an unnecessary comment, and that I felt more at home with ChatGPT’s replies. Both reactions were genuine. August 28 post / October 24 post

That is a personal preference from everyday work. We could not measure Claude in the guest test, and ChatGPT also added an explanation when given the broader editing brief. Alongside accuracy I would score one more thing: could I use the reply as delivered, without having to trim it?