You are 90 messages into a project and your AI assistant starts forgetting things you told it an hour ago. It repeats itself. It gets sloppy. You did nothing wrong. The chat just got too long, and there is a one minute fix that does not mean starting from zero.
Ben Gusberg, an AI startup founder, shares this fix in a short video. Watch it first, then come back for the full setup, a copy and paste prompt, and the limits nobody mentions.
Prefer to watch on YouTube? Open the video here.
Why long chats go bad
Picture a desk. Every message you send and every reply you get stays on it. After a while, the one sticky note that matters is buried under 60 others.
Researchers call this context rot. A July 2025 technical report from Chroma tested 18 leading models and found that performance drifts as the input grows, even on simple tasks. Anthropic’s own context window documentation says the same thing in plain terms: as the token count grows, accuracy and recall fall.
Here’s the thing. A bigger context window does not make this go away. It just gives the problem more room to grow.
About that 39% number
The video mentions a drop of up to 39% in reasoning quality. That figure is real, but it is easy to misread, so let’s break it down.
It comes from a paper by Philippe Laban, Hiroaki Hayashi and colleagues, LLMs Get Lost In Multi-Turn Conversation, later accepted at ICLR 2026. The team tested 15 models on six tasks. They compared giving a model the full instruction at once with feeding it the same instruction piece by piece across several turns.
| Setup | Average performance |
|---|---|
| Full instruction in one message | About 90% |
| Same instruction spread across turns | About 65% |
That is the 39% average drop, and it showed up even in two turn chats. The more useful finding is why. Models made early assumptions, leaned on them, and did not recover after a wrong turn. The loss was mostly in reliability, not raw ability.
This is a different test from Chroma’s, but both point the same way: long, winding conversations make answers worse. It also explains why a clean restart helps so much. A fresh chat has no wrong turns in it.
The fix: a handoff summary you control
Ben’s idea is simple. You teach your assistant one trigger phrase, /context transfer. When you type it, the assistant writes one tidy summary of your project. You paste that summary into a brand new chat and carry on.
Anthropic describes the same move as compaction in its guide on effective context engineering: summarize a conversation that is nearing its limit, then start again with that summary. The difference here is that you read the summary before it goes anywhere. What this really means is that you decide what survives.
Set it up in 2 minutes
First, copy this prompt. It is kept short on purpose, so it fits inside ChatGPT’s free plan limit.
Handoff rule: when I type /context transfer, stop and write one handoff summary I can paste into a new chat.
Use these headings:
Goal: what I am trying to achieve, in two sentences.
Decisions: choices we locked in, and why.
Key facts: names, numbers, dates, constraints.
People: who is involved and their role.
Links and files: every link or file name mentioned.
Done: what is finished.
Open tasks: what is left, in priority order.
Style rules: tone, format, things to avoid.
Keep it under 400 words. Write it so someone with zero history could continue the work. Only include what we actually discussed. If something is unclear, write Unclear instead of guessing.In Claude
- Open Settings.
- Find the Instructions for Claude field.
- Paste the prompt and save.
Instructions here apply to all your chats. If you only want the trigger inside one piece of work, paste it into that Project’s instructions instead. Anthropic explains both options on its personalization help page.
In ChatGPT
- Click your profile picture and open Settings.
- Choose Customize ChatGPT or Personalization. The label depends on your version.
- Open Custom Instructions, paste the prompt, and save.
OpenAI’s custom instructions help page says Free and Go users can save up to 1,500 characters, and that changes apply to all chats right away.
How to use it when a chat starts to slip
- Spot the signs. Details you gave earlier come back wrong, answers get vaguer, or the assistant repeats itself.
- Type
/context transfer. Send only that. Wait for the summary. - Read it for 30 seconds. Fix anything wrong. Add anything missing. This step is the whole point.
- Open a new chat. Start with this line, then paste the summary: Here is a handoff from a previous chat. Read it, tell me in three lines what you understand, then wait for my next instruction.
- Check the three lines. If they match your project, keep going. If not, correct it now, while it is cheap.
Here is a short example of what you might get back for a freelancer working on a client proposal.
Goal: Finish a proposal for a 3 month website redesign for a local bakery chain.
Decisions: Fixed price, not hourly. Two design rounds included. Payment split 50/50.
Key facts: Budget ceiling is 4,000 USD. Launch target is the first week of December.
Open tasks: Write the timeline section. Draft a polite note about extra revision costs.
Style rules: Plain English. No jargon. Short paragraphs.
That is the full project in a few lines. Your new chat starts sharp instead of starting over.
Signs it is time to reset
- The assistant forgets a rule you set at the start.
- It contradicts something it said earlier.
- You have gone back and forth correcting the same mistake.
- The chat has dozens of messages, pasted documents, or long tool outputs.
- You changed direction twice and the early messages no longer apply.
One more habit worth adopting: if you change your mind midway, do not just keep arguing in the same thread. Run the transfer and start clean. Given that models struggle to recover from a wrong turn, this is often faster than fixing it.
Honest limitations
- It is not a real command.
/context transferis just a phrase your instruction teaches the assistant. It usually works, but a very long, rotted chat may follow it loosely. - Summaries lose detail. Exact wording, tone, and long drafts get squeezed. If a draft matters, save it separately and paste it into the new chat in full.
- Files do not travel. The summary lists file names, but the new chat cannot see them. Upload them again.
- Errors can get copied. If the assistant misunderstood something earlier, the summary may repeat it. That is why the 30 second read matters.
- The studies are not about every chat. Short conversations are generally fine. This habit pays off on long, multi step projects.
- Cross check big answers. A fresh chat helps, but it is not a guarantee of accuracy. Our Cuey review covers a free way to compare answers across models.
Try it today
- Paste the prompt into your instructions. It takes two minutes.
- Pick your longest running chat and type
/context transfer. - Paste the result into a new chat and compare the quality of the next answer.
Credit for the tip goes to Ben Gusberg. The prompt above is my own version, so tweak the headings to fit your work.
Know someone who loses their mind when ChatGPT forgets things mid project? Send them this post and save the prompt for your next big project.
Sources
- Ben Gusberg, video on solving context rot (YouTube)
- Hong, Troynikov and Huber, Context Rot: How Increasing Input Tokens Impacts LLM Performance (Chroma)
- Laban, Hayashi and others, LLMs Get Lost In Multi-Turn Conversation (arXiv)
- ICLR 2026 listing for the same paper
- Anthropic, Context windows (Claude documentation)
- Anthropic, Effective context engineering for AI agents
- Claude Help Center, Understanding Claude’s personalization features
- OpenAI Help Center, ChatGPT Custom Instructions
Leave a Reply