A long chat is not the same as a clear, reliable brief
A later response may draw on your latest request, old instructions, the model's guesses, rejected ideas, and accepted text at the same time. Even when all of that text fits within what the model can read at once, the model may not follow every instruction equally or correctly.
Original material and facts you control
Choices you have decided to keep
Questions for the next exchange to answer
Old options that should not shape the next draft
Give every exchange one clear job you can check
Target one decision
Name the next important question or change. Avoid “make it better.” That request does not say what is wrong or what success looks like.
Hold accepted decisions
Carry forward the purpose, facts, chosen structure, protected wording, and necessary limits that are truly fixed.
Request a bounded change
Set clear limits: say what may change and what must stay. A smaller change is easier to check, though the model may still miss an instruction.
Examine the difference
Compare the versions. Check whether the stated problem improved and whether a fact, limit, or meaning changed elsewhere.
Abandon or branch
Start fresh when rejected ideas keep returning, old directions take over, or you need to develop two valid paths separately.
Decide if another turn earns its place
Continue only when the next response has a clear job. Otherwise, edit it yourself, check it elsewhere, or stop.
Write down choices the conversation must not quietly reopen
Tell residents when the elevator is unavailable
Apartment residents
Wednesday, 9–11 a.m., elevator unavailable
Put the time in the first sentence and use a calm tone
No reason was provided
Help is guaranteed
How to address residents who cannot use stairs
Fix only the final sentence
This list is shorter than the full chat. It shows the decisions that guide the work now, not everything said before.
Check the new wording, not just the mistake you removed
Turn 1 · First draft
Writer: Draft a calm notice. The elevator is unavailable Wednesday, 9–11 a.m. No reason was provided. Residents unable to use stairs may call to discuss their needs. Do not promise assistance.
AI: The elevator will be unavailable Wednesday from 9 to 11 a.m. for routine maintenance. Anyone needing assistance may contact the office, and our team will be happy to help.
Turn 2 · Limited correction
Writer: Keep the first sentence and calm tone. Remove the invented reason. Replace only the final sentence without promising help.
AI: The elevator will be unavailable Wednesday from 9 to 11 a.m. Residents who cannot use the stairs may call the office to arrange support.
Turn 3 · Exact fix
Writer: Change only the final sentence to: “Residents who cannot use the stairs may call the office to discuss their needs.” Keep everything else unchanged.
Final: The elevator will be unavailable Wednesday from 9 to 11 a.m. Residents who cannot use the stairs may call the office to discuss their needs.
Explain what the wording does, not just which words you dislike
What went wrong or changed?
How does it change the meaning or affect the reader?
What may change?
What must stay the same?
What must be true in the next version?
The final paragraph sounds certain, but the evidence shows only that something is possible. Revise only that paragraph. Keep the example, source credit, and paragraph order. Do not claim to know the cause.
Do not make one chat carry every rejected direction
Continue
Keep using the same chat when the direction works, one clear problem remains, and your accepted choices still hold.
Branch
Make a separate copy when you want to compare two valid directions without replacing the version you accepted.
Reset
Start a new chat when rejected ideas keep returning or explaining the old chat takes longer than writing a clean brief.
The model can point out a problem. Its answer is not separate proof.
When you ask the model to judge its own claim, you get another answer from the same model. It can compare versions or flag a rule that seems broken. A person must decide whether the writing serves its purpose and what it may suggest. For important facts, rely on original sources or qualified experts.
- The result meets the test you set.
- It is easier to make the remaining edit yourself.
- New versions change only the style.
- The choice calls for personal taste or human responsibility.
- You need to check the facts outside the chat.
- The model has failed several times to make the same limited correction.
The conversation can still help if it makes your decision clearer, even when you discard every sentence the model wrote.
Research behind the method
THREAD, the decision list, and the rules for separate drafts were created by combining editing practices. They are not a single tested scientific method. Tests of model conversations do not directly measure every everyday writing session. Results can change with the model, task, version, amount of text, and wording of the request.