OnboardMission Path
In previous lessons you learned the model forgets things that DID happen in the conversation. This lesson is the mirror: sometimes the model invents things that NEVER happened, because you gave it cues that suggest they did.
The model predicts the next token. If your message starts with "as we said" or "about the file I sent you", the most probable words afterwards are those of a prior agreement or a file reference. Even if no agreement and no file exist.
The model won't tell you "I don't have that context". It will fabricate the missing context and continue as if it had it.
On the right you have the three classic cases:
| Dangerous habit | Safe habit |
|---|---|
| "Go ahead, as we discussed." | "I want to do X. Plan: A, B, C." |
| "The file I sent you." | (paste the file content) |
| "You know my setup." | "My setup is: X, Y, Z." |
Golden rule: every new conversation starts from zero. The model doesn't "remember" yesterday's conversation, the file you opened, or the project you're working on. If you want it to know, you paste it.
This gets subtler within ONE long conversation. The model "remembers" the first turns, but with less weight already. If after 20 turns you make an ambiguous reference to something from turn 2, the model can invent the connection instead of checking.
Look at the three cases, identify the trap in each. The next time you're about to write "as we discussed" in a first turn, you'll have the reflex to spell out the context instead.