Know when to start a fresh chat
Yesterday's task was a friendly invitation. Today's is a factual briefing. If the new answer keeps sounding like an invitation, the old conversation may still be steering it.
What you'll make
a short, checked handover and a clean brief that starts the new job in a fresh chat, without the old job’s tone, dates or rejected ideas.
What you'll need
ChatGPT, Claude, Gemini or Copilot in your work account, the chat you’ve been working in and about ten minutes.
Example result.
New task: note comparing two support rota options for Priya (operations manager).
Use only: the pasted descriptions of Option A and Option B. Neither is agreed.
Keep unresolved: November staffing numbers.
Do not import: the team-lunch invitation's tone, date or sign-up link.1. Decide whether the job is still the same
Stay in the current chat when you’re refining the same output, from the same sources, for the same reader. The useful context is already there.
Start fresh when any of these happen:
- The job changes. New purpose, new reader, new source, or an unrelated subject.
- Two tasks must not borrow each other’s facts. Different clients, different teams, different confidentiality.
- The chat starts drifting. The assistant repeats itself, forgets an instruction you gave earlier, or brings back a version you rejected. That’s a reason to hand over and restart, even for the same job.
A quick test: write one sentence describing what you need now. If it no longer matches the original request, a fresh brief will be clearer than another “actually…” message.
Why this works: a long thread carries every earlier request, including the ones you’ve since changed your mind about. The model can’t always tell which ones still count.
What you should see: a one-sentence description of the new job, and a clear decision to stay or restart.
2. Ask for a handover, then check it
In the old chat, ask for a short handover rather than copying the whole conversation, which may include rejected drafts or details the new task doesn’t need.
Prepare a short handover for a new conversation about [new task]. Include only the facts that remain relevant, accepted decisions, unresolved questions and the sources they came from. Label proposals as proposals. Exclude rejected draft wording and unrelated details. Do not add facts or treat your earlier suggestions as decisions.[new task] is your one-sentence description from step 1.
Why this works: asking for decisions, proposals and open questions separately stops a suggestion from quietly becoming “agreed”.
What you should see: a handover of a few lines. Check it against the original material and delete anything the new job doesn’t need. A summary can repeat an earlier mistake, so read it before you paste it.
3. Start a new chat with a complete brief
Open a fresh conversation:
| Tool | Start fresh | Also useful |
|---|---|---|
| ChatGPT | New chat in the sidebar | On the web, ⋯ › Branch in new chat on a message keeps the original intact |
| Claude | New chat | Long chats use more of your usage limit |
| Gemini | New chat, top left (open Menu first if hidden) | Temporary chat sits next to it |
| Copilot | New chat in the left navigation | — |
If the work belongs in a project or workspace, open that one and check its standing instructions. A new chat doesn’t necessarily clear memory, project instructions or connected access; those are separate settings.
New task: [job].
Reader: [audience].
Purpose: [what the reader should understand or decide].
Use only these checked facts and sources: [context].
Output: a short note, under 200 words.
Do not import an earlier conversation's tone, dates, decisions or rejected suggestions. If an important fact is missing, say so instead of guessing.[job] is the new task, [audience] the reader, [what the reader should understand or decide] the purpose, and [context] your checked handover. Change the output line to suit the job.
Why this works: the brief states the job from scratch, so the model isn’t inferring it from a history of edits.
What you should see: an answer shaped for the new reader, using only the facts you pasted.
Check the result
- Look for an old date, audience, tone or source that isn’t in the new brief.
- Check every fact against your handover. Anything new needs a source.
- Make sure open questions stay open rather than being filled in.
The usual failure: a plausible detail from the old task slips in: a date, a friendly sign-off, a link. It isn’t nonsense; it’s a real fact from the wrong job, which is why it’s easy to miss. Read the answer asking “where did this come from?” for every specific.
The answer includes [old detail], which is not in this task's checked context. Remove it. Rebuild the affected passage using only the facts in the current brief, keep unresolved information visible and identify any fact you cannot support.[old detail] is the phrase that came from the earlier task. If the same detail keeps returning across new chats, check memory and project instructions in your tool’s settings; opening another chat won’t change them.
Worked example: from a lunch invitation to a rota note
Fictional example. Earlier in the week, a chat drafted a cheerful invitation to a team lunch on Friday 23 October, with a sign-up link. Later, in the same chat, the team lead asked for a note to Priya comparing two support rota options.
Option A: fixed shifts. Same two people on 08:00–16:00 and 11:00–19:00 each week.
Option B: rotating shifts. Everyone covers one late week (11:00–19:00) in four.
Neither option is agreed. November staffing numbers are not confirmed.The answer, still in the old chat:
Hi Priya! Two options for the new rota. Option A keeps the same people on early and late shifts. Option B shares the late shift, one week in four. Join us next Friday to launch the new rota!
- The failure. The bold line comes from the lunch invitation: its tone, its Friday and an event nobody planned. The note also skips the unconfirmed staffing.
- The repair. A fresh chat with the brief from step 3, using the handover in the example result, gave: “Priya, two rota options for a decision. A: fixed early and late shifts, same people each week. B: everyone covers one late week in four. Staffing for November isn’t confirmed, so neither option’s cover can be checked yet.”
- Why it worked. The new chat had no lunch to borrow from, and the brief named staffing as unresolved, so the answer said so.
Task kit for this job: Download (ZIP) · Reusable instructions
Go deeper
Optional training: Keep a reusable task record
For work that runs over weeks, keep the purpose, accepted decisions, sources and open items in a document outside the chat. Update it when the work changes. When you restart, paste the record instead of asking the assistant to summarise a long thread. It’s a checked starting point that doesn’t depend on the model remembering everything correctly.
Optional training: Separate exploring from doing
One chat can explore several options freely. When you move to the doing (drafting the final note, sending the email, writing the plan), start a new chat that states which option was chosen, who agreed it and what the output is for. Rejected ideas from the exploring chat shouldn’t become a hidden backlog in the final one.
Optional training: Fix the right layer when the problem keeps coming back
If a style or fact follows you into every new chat, it’s probably stored above the conversation. A recurring style preference points to custom instructions or memory. A fact that only belongs to one project points to that project’s files or instructions. An answer drawing on the wrong documents points to connector or source settings. What your assistant remembers covers those controls.