Stop explaining the same thing. Give AI a clearer place to start.
When a useful conversation drifts, another long correction may make it harder to follow. Check the source, the task and the decisions that still apply before asking for more work.
By Joe McKenna. Adapted from his September 29, 2026 workshop.
A useful restart carries forward the source and the decisions you checked, with a clear statement of what you want next.
You told it already. Why is the answer different?
You have been working with AI for a while. The first answer was useful. You changed a few things, tried another idea, asked for a different format, and now it has brought back something you already corrected. You are spending your time explaining the project again.
In the workshop, I compared that to going to college with the same book and notepad for every class. Eventually, the notes stop helping you tell one subject from another. The comparison is about organization: the material may be there, but the job you want done is getting harder to see.
My business example was a familiar frustration. You want a particular contact address on a piece of marketing, but AI uses a different one it found elsewhere. That can happen even when the alternative looks perfectly reasonable. Your preferred source and the rule for using it need to be clear.
Before adding another paragraph of correction, ask a smaller question: what is this answer actually working from?
Why one long conversation gets tangled
What to notice: Notice Joe’s college-book and notepad analogy. A source collection helps organize context; it does not guarantee complete recall or correct answers. The percentage and memory explanation in the live talk are conversational shorthand, not measured usage data or a complete technical account.
1:09. From The One-Person AI Business, 14:27.00–15:35.56.
Visual context: Joe explains beside the workshop lesson and Gemini browser screens. The college-book analogy is spoken; the interface supplies context.
Read the clip transcript
What you want to be able to do is actually use AI whereas accessing a knowledge base that is bigger than the chat you're on, okay? So this is what 99% of most people using AI have. So here, if you see my screen, I'm in a chat right now on Gemini and I'm having a conversation here. I say this, it says this, and we go back and forth. But this entire conversation is all it has to decide what the context is, how it's gonna answer to me, what the rules are. And as this conversation goes on, it starts to lose the ability to remember everything above it. It would be like going to college and every class you go to, you're using the same book and the same notepad paper. At some point, it's like doesn't make any sense anymore. What's what? And so AI gets super confused, but this is the depth that you can have a conversation in normal AI is at the chat level. That's this, okay? So if I back up here, we need to go past that where you actually have a knowledge base that's tapping in and out of, coming in and out of it.
Automatically transcribed from this excerpt and aligned locally. Transcript accuracy and word synchronization still require a human listening review.
The chat on your screen is only part of the picture
Context is the information available to help produce the current answer: your request, relevant conversation, instructions and material the tool can access. A context window limits how much information a model can work with together. Google warns that very large uploads can lead Gemini to miss connections or details. Read Google's explanation of context and large files.
That does not mean every bad answer is a memory problem. The file might be missing. Two instructions might conflict. An old source might still be selected. The answer may simply be wrong. A longer chat or a larger model does not settle those questions for you.
Some products can bring in saved information or retrieve passages from a collection of sources. Retrieval means finding material to use for a particular question. Having a file in a collection is different from checking that the answer used the right passage. That is why I want evidence of the source, not just a confident statement that everything is understood.
Find the break before you restart
Use these checks in order. You may only need to fix one thing.
- Is the intended source available? Look for the actual attachment or selected source. Ask for one specific detail and its supporting passage. A filename alone does not show that the contents were read.
- Are there competing versions? Name the version to use. If a newer date appears in a separate correction, say that it replaces the earlier date.
- Has the task changed? A draft for a colleague, an announcement for students and a public flyer have different needs. Tell AI which job it is doing now.
- Is the answer adding its own material? Ask it to separate source facts from suggestions. If the source does not establish a detail, leave that detail unresolved.
If the answer is wrong with the correct material plainly available, treat it as an error to correct and verify. Do not keep adding background in the hope that sheer volume will fix it.
Carry the right details into the next attempt
There are three problems to fix: the invented day, the superseded room and the changed preparation. Asking AI to make the announcement better leaves those problems unnamed. Give it the current facts and explain which earlier detail has changed.
We are preparing a short announcement for students. Current approved facts: - Peer review is Thursday at 2 p.m. - The room is now 210. This replaces Room 204. - Students bring one printed draft, not a final paper. Use these facts for this task. Do not add a date or other requirements. Write a two-sentence announcement. Then list the facts you used so I can check them.
A suitable authored example would be: Our peer-review session is Thursday at 2 p.m. in Room 210. Please bring one printed draft.
This is an illustration of the check, not a tested AI response. Compare each detail yourself. The restart succeeds when the confirmed facts survive the change in wording.
Keep a short handoff instead of the whole conversation
When a chat has collected several abandoned directions, a new task can be easier to manage. Before leaving, save a short handoff that you can inspect. Include the goal, audience, source versions, approved decisions and unresolved questions. Exclude ideas you have already rejected.
You can ask AI to draft that handoff, but read it before reusing it. Otherwise an unnoticed mistake becomes the starting instruction for the next conversation. Keep the original source available alongside the handoff; a summary can omit an exception you need later.
For the announcement, the handoff is only a few lines. For a longer project, it might identify the current procedure, explain the audience, and say which decisions still need the document owner's answer. Its purpose is to make the next task clear, not to preserve every message.
If you keep repeating the same source setup, the next step is a reusable collection of sources. That article covers organizing and maintaining the collection. Here, the job is to diagnose why this particular answer drifted.
Try one small reset you can check
Choose a conversation that has become frustrating. Write down one thing it got wrong. Identify the source or decision that settles it, then give AI that material with one small task.
Check the result against the original. Did it keep the right names, dates and requirements? Did it drop an exception? Did it add something you never supplied? If it still misses a detail, narrow the task to that detail before asking for a complete draft.
You do not have to become an expert in how a model works to do this. You need a clear starting point and an answer small enough to inspect. If the source itself is unclear, return to the person responsible for it. AI cannot supply an approval that nobody has given.
Questions that come up
Should I always start a new chat?
No. A short correction may be enough when the task and source are still clear. A new chat is useful when several competing directions make the current task hard to follow. Carry over a checked handoff and the relevant source.
Does a notebook make AI remember everything?
No. A notebook can organize reusable context, but you still need to check source access, versions and the answer itself. Product settings and features affect what is available.
What if the source is attached but AI still gets it wrong?
Ask it to locate the specific passage and compare that passage with the claim. Correct the error directly. For a long source, try the relevant section with its surrounding context. Escalate unresolved questions to the source owner.
Can I use AI to summarize the old conversation?
Yes, as a draft handoff. Check that it preserves approved decisions, labels open questions and removes rejected ideas. Keep the original source too; a summary is not a complete replacement.
15 Days of AI
Put this into practice with Joe.
Give AI useful context. Work with your own materials. Create something you can use. Explore the independent classes and choose the ones that fit your goals.
Explore the classesAbout this lesson
First-person teaching is adapted from Joe McKenna's September 29, 2026 workshop and the automatic timestamped transcript. The college book/notepad comparison and wrong business-contact example are source-derived paraphrases. Context definitions, diagnostic sequence, handoff method, fictional course announcement, prompt and FAQs are editorial additions. No direct quotation, continuous viewing, audio verification or tested AI output is claimed. Workshop claims that ordinary chats only use the visible conversation and that extra context prevents errors have been narrowed. Product documentation checked September 30, 2026.
Source passages: 14:01–17:51; 39:19–41:13. Source: The One-Person AI Business, September 29, 2026.
Official guidance checked September 30, 2026. Product features can change.
AI-assisted organization and supplementary examples. Prepared from Joe McKenna’s workshop.