When to restart and when to keep going
Restart when the chat is no longer a reliable record of the work. Signs include a standing rule that lapses, a fact that changes between turns, an answer that suddenly grows several times longer, or a conversation that has moved away from its goal. Keep going while the task is short, the rules still hold and each answer builds on the last one.
The Context Drift Heatmap can show which checkable rules lapsed before you decide. Vendors describe a related trade-off. Anthropic’s documentation says that “response quality degrades as a conversation grows,” and OpenAI describes its compaction as helping “balance quality, cost, and latency as conversations grow.” (Anthropic, Compaction; OpenAI, Compaction) Both are API features that act on the provider side as a conversation grows. Anthropic’s replaces older turns with a summary; OpenAI’s runs when the token count crosses a threshold you set. Starting a new chat yourself is a different action, and the sources do not compare the two.
| Situation | Usually | Why |
|---|---|---|
| Rules still hold and the task is in progress | Continue | The transcript is still doing useful work |
| A standing rule has lapsed more than once | Restart with a handoff | Repeating the rule has not held it |
| Facts changed between turns | Fix the facts, then restart | The newest value and its source belong in the handoff |
| One answer is several times longer than the rest | Read it first, then decide | It may repeat earlier text, and reading it costs less than restarting |
| The chat has moved to a new goal | Restart | The old transcript carries the wrong goal |
The seven parts of a handoff
A handoff is one short message with seven labelled parts. Put every part in the same message. The Laban et al. simulation found that “providing all the information at once (1-shard) is the only effective method to improve reliability.” That result describes a simulated single-message setting. A handoff is one message by design, but that study did not test handoffs.
| Part | What goes in it | Invented example |
|---|---|---|
| 1. Goal | One sentence on the outcome | A 900-word spring newsletter for a fictional bakery |
| 2. Decisions | Choices already made, with the reason | British English, because the bakery is in Leeds |
| 3. Standing rules | Rules that must hold in every reply, written to be checked | Under 120 words per section; no emojis; cite each fact as [n] |
| 4. Open questions | What is still undecided, with the options | Opening date: 4 April or 11 April? |
| 5. Key facts with sources | Facts you checked, with the source and date | Opening hours Mon to Sat, 8:00 to 16:00 (owner email, 3 October) |
| 6. Next step | The single task for the first reply | Draft the opening paragraph using the rules above |
| 7. Output format | The shape of the reply | Headings, then a bulleted list of changes |
A copyable handoff template
Paste this at the top of a new chat, fill in each bracket and delete any empty part. Keep the note short enough to read in one pass. Attach the sources you cite, not the full old transcript.
## Handoff: [short project name]
**1. Goal**
[One sentence: the outcome I want from this chat.]
**2. Decisions so far**
- [Decision] (reason: [why])
**3. Standing rules (apply to every reply)**
- [Rule, written so it can be checked, e.g. "under 120 words"]
**4. Open questions**
- [Question] (options: [A] / [B])
**5. Key facts, with sources**
- [Fact] (source: [URL or document name], checked [date])
**6. Next step**
[The single task for the first reply.]
**7. Output format**
[e.g. headings, then a bulleted list; no tables]
Before you answer: restate sections 3 and 4 in one line each, then do the next step.
What to carry over and what to leave behind
Carry over what the next answer needs in order to be right. Leave behind what only mattered while the answer was being worked out. The split is simple, but it is easy to get wrong when a long thread feels important.
- Carry: decisions with their reasons, rules written so they can be checked, facts with their sources and dates, open questions with options, and the one next step.
- Leave: the full transcript, dead-end drafts unless you record why they failed, old hedges that are no longer true, and large pasted files, which you can replace with a summary and a link.
Carrying the handoff matters more because chat apps may already drop older turns. Anthropic’s documentation notes that claude.ai “can also manage the context window on a rolling ‘first in, first out’ basis.” (Anthropic, Context windows) Summaries also lose detail. Anthropic’s engineering team warns that “overly aggressive compaction can result in the loss of subtle but critical context.” (Anthropic, Effective context engineering) A handoff is a summary you write, so check it against the old chat for numbers, exceptions and dates.
Test the first reply before you rely on it
Ask the new chat to restate sections 3 and 4 before it does any work, as the last line of the template does. If a rule is missing or changed, fix the handoff and start again rather than correcting it mid-conversation, because the wrong version stays in the thread.
Then compare the first answer with your rules. The drift guide’s checks show how to run a restatement, a re-ask and a length check. A canary that passes does not prove that later answers will follow the rules, so keep checking.
Before you write the handoff, run the old chat through the Context Drift Heatmap. Its ledger shows which rules lapsed and at which turn, which tells you which rules need clearer wording in section 3. To see how a chat fills a window size, the Context Window Visualizer compares a multi-turn chat with 8K to 200K windows.
What the research does and does not show about restarting
None of the research cited here tests a restart with a summary. Laban et al. covered recap and snowball conditions, both of which restate earlier information within one conversation. The paper calls Recap “an unrealistic setting because the intervention is conducted on the last turn of the conversation.” Snowball, which repeats earlier user turns, gave gains the authors describe as “realistic.” It can “mitigate the Full-to-Sharded performance deterioration by 15-20%,” in two models and four tasks. Snowball is a different method from a restart, and its gain was measured in simulation. (Laban et al., LLMs Get Lost In Multi-Turn Conversation)
The narrow claim is this. In that study, giving everything in one message was the most reliable setting, and a recap on the last turn was not a realistic test. A handoff tries to approximate a single full message in a new chat. Whether it matches the original is not something this research measured.
Anthropic’s engineering post describes summarizing and restarting as a practice it uses, and it gives no measured effect. It defines compaction as “the practice of taking a conversation nearing the context window limit, summarizing its contents, and reinitiating a new context window with the summary.”
Common mistakes, and what this cannot tell you
- Pasting the whole transcript under the handoff. That brings back the growth you were trying to escape.
- Writing rules that cannot be checked. “Be concise” is hard to verify. “Under 120 words” is not.
- Trusting the summary without checking numbers. Names, dates and exceptions are common places for a summary to lose detail.
- Leaving open questions out. The new chat will then settle them without telling you.
- Calling the handoff a fix. It is a clean start. Its effect on your task is not measured here.
This guide cannot tell you how a given app stores, trims or summarizes chat history. For that, read the provider’s documentation for the product you use.
Common questions
How long should a handoff be?
Short enough to read in one pass, which is often a page or less. None of the sources gives an ideal length. Keep what the next answer needs and cut anything it would not use.
Should I ask the AI to summarize the chat for me?
You can, and it may save time. Check the summary against the old chat for numbers, dates and exceptions. Anthropic advises maximizing recall in a summary, and suggests writing your own summarization prompt when the default summary drops something later turns need.
Is a new chat better than compaction?
The sources cited here do not compare them, so no study answers that. Anthropic’s compaction is an API feature that replaces older turns with a summary. OpenAI’s is a server-side feature triggered by a token threshold. A new chat is a choice you make by hand. Use whichever your tool supports, and check the first answer either way.
What if the new chat ignores the handoff?
Ask it to restate sections 3 and 4, then check the reply against them. If a rule is vague, rewrite it so it can be checked, such as a word limit, and start again. A chat that ignores a rule once may ignore it again, so test the next answer too.
Does the handoff work in any AI tool?
The template is plain text, so you can paste it into most chat tools. The studies cited here tested named models between 2023 and 2025, so they do not predict how another model will respond to a handoff.
How do I know when a chat is too long?
The sources give no fixed turn count. Use the signs in this guide instead: a lapsed rule, a changed fact or a sudden jump in answer length. A token count can help, but it is an estimate.
Sources and further reading
- Laban et al., LLMs Get Lost In Multi-Turn Conversation, arXiv, May 2025
- Anthropic, Effective context engineering for AI agents, Sep 2025
- Anthropic, Context windows, Claude Platform docs
- Anthropic, Compaction, Claude Platform docs
- OpenAI, Compaction, API docs