When it goes wrong: seven fixes to try first.

When an assistant starts forgetting things, repeating itself or sounding sure about something wrong, the instinct is to blame the model and try a bigger one. Most of the time the model is fine. The problem is what it was given, and the fix takes minutes.

Start from the symptom. Each one points to a likely cause:

Forgets what you told it

The chat has got long and your early instructions are crowded out. Fix one.

Long, muddled answers

It was asked for too much at once. Fix two.

Wrong facts from a document

It can't read the file properly, or the file is pasted into the chat. Fixes three and four.

Confident answers it made up

Nobody told it what to do when it doesn't know. Fix five.

Good at one thing, odd at another

It's doing too many jobs. Fix six.

Worked yesterday, fails at one step

A connection changed. Fix seven.

Course one called the context the desk: everything the model can see at once. A long conversation fills the desk, and the instructions you gave at the start get pushed out of view. That's when it starts forgetting the tone you asked for, or the thing you said not to do.

Start a new chat with a short recap of what matters. Keep one chat per job or project, rather than one long thread for a week of different tasks.

We're drafting replies to tenants for a lettings agency. Friendly, plain English, under 150 words. Never quote a fee. Here's the message to reply to:

Don't ask for the summary, the email and the spreadsheet in one go. Ask for one, check it, then ask for the next. Each step gets the model's full attention, and when something goes wrong you can see which step it was.

Then check that it can actually read what you gave it. A scanned PDF can be just a picture of words. Some assistants can read images, but a clean copy with real text is far more reliable. A quick test: if you can't select and copy the text in the file yourself, export or ask for a text version.

Pasting a forty-page guide into a chat puts all forty pages on the desk, every time. It crowds out everything else, including your question.

Store it as knowledge instead. Then the assistant looks up the part it needs and answers from that, the way course one's RAG lesson described. Most assistant tools have a place for this: knowledge, files or a project library.

A model predicts the next piece of text. If nothing tells it what to do when it doesn't know, it fills the gap with something that sounds right. One line in its instructions changes that.

If you don't know the answer, or it isn't in the documents you've been given, say so and ask me. Never guess a date, an amount or a name.

An assistant set up to answer tenants shouldn't also write the social posts. Different jobs need different instructions, tone and documents, and one set of instructions trying to cover both does neither well. Make a second assistant.

And when you fix a connection, such as reconnecting an email account or an app, run one small, harmless test before trusting it with real work. Often the model answered perfectly and the failure was in the step that sends or saves the result.