GeminiPrompting mistakesLevel 1: Basic prompts

Why Does Gemini Get Worse in a Long Chat?

A long chat is more work for Gemini to reread each time, and on smaller plans it can run past what the model can hold. Google says exceeding the context window can cause responses to miss details, and that long histories use more of your limit. Start a new chat for each topic, and carry over a short summary.

Prompt to copy
Before I start a new chat, write a handover I can paste as my first message:

- Goal: what I'm trying to achieve
- Decisions made, and why
- Rules I gave you that still apply
- Current state of the work, with the latest version in full
- Next step

Keep it under [300] words, excluding the latest version. Leave out anything I rejected, and don't add anything I didn't say.
Do
  • New chat for each new topic
  • Carry over a short summary
  • Save lasting rules in a skill
Don't
  • One chat for a whole project
  • Stack corrections on a bad draft
  • Change topic mid-chat
Rule of thumb, not a Google chart, based on its advice to start a new chat for each topic.

Try it now

  1. 1Open your longest active chat and paste the prompt above.
  2. 2Read the handover, and correct anything it got wrong or left out. That also shows what it had lost track of.
  3. 3Click New chat, paste the handover, and continue the work.
  4. 4You'll know it worked when Gemini follows your rules again without being reminded.

Use it when

When Gemini repeats a rule you already changed, forgets an early decision or starts mixing up two topics. Change chats before that point on a free account, especially if you've attached a large file.

Skip it when

A chat that's short and still on topic. There's no reason to restart it, and you'd lose context you still need. Also, if a project needs its files and instructions kept between chats, a new chat alone won't do it, so use a notebook or a skill.

Why it works

The chat history is part of what Gemini has to process with each message. Google's limits page gives the app's context windows as 32k tokens with no plan, 128k on AI Plus and 1 million on Pro and Ultra. Using Google's figure of 60 to 80 English words per 100 tokens, 32k is about 19,000 to 26,000 words (arithmetic, not a Google figure), and any file you attach takes up that same room. Also, if you've hit a usage limit, Google says subscribers continue on Flash-Lite, so a chat can get weaker partway through for a reason unrelated to its length.

Example

Say you've spent a week in one chat on a grant proposal, with the budget, the narrative and a side question about your visa all in there. On your last edit, Gemini used a budget figure you changed days ago. Rather than correct it again, ask for the handover and check the figure there. Then open a new chat for the narrative, and another for the visa question.

Going further

Put things that outlast a chat somewhere built for them. A rule you repeat belongs in a skill, which you can call from any chat. A set of files and instructions you keep coming back to belongs in a notebook in the Gemini sidebar, with its own instructions under Notebook settings. And if Gemini seems to be stuck on a bad direction, Edit text on an earlier prompt regenerates from that point, rather than adding another correction on top.

Common mistakes

  • Correcting the same mistake three times in one chat. Each correction adds to the history without removing the error.
  • Using one chat for unrelated topics. Google recommends a new chat for each topic.
  • Blaming the length when you've hit a usage limit. Click More > Show thinking steps to see which model answered.
  • Trusting the handover without reading it. If a detail's already lost, the summary locks that in.

Sources

Checked 6 November 2026. Products change often, so confirm details with the vendor.

Related

Beyond the course