Conversations have a shape

Long threads go bad in predictable ways

A conversation is not a relationship with an assistant that gets to know you. It is a growing block of text that the model re-reads in full before every reply. That single fact explains everything that goes wrong in long sessions.

Drift. Forty messages in, the model is conditioning on forty messages of context, including three abandoned approaches and a tangent about something else. Its answers start blending them. You get an essay plan that quietly reverts to a structure you rejected twenty minutes ago.

Anchoring. Once a wrong claim is in the context, it tends to persist. The model treats its own earlier output as established. Corrections help less than starting again, because the wrong version is still sitting there being conditioned on.

Sycophancy compounding. Assistants are tuned to be agreeable. Over a long thread of you pushing back, that tuning accumulates, and you end up with a model that agrees with everything you say. At that point it has stopped being useful as a critic, which is usually why you opened it.

Silent truncation. When the thread exceeds the window, the oldest content is dropped or compressed. Nothing tells you. Constraints you set at the start stop applying.

The fix is cheap and unglamorous

Start a new chat far more often than feels natural. New subject, new chat. Wrong turn, new chat. Twenty messages in, new chat. Copy forward only the two or three things that matter: the current draft, the constraint list, the conclusion you reached. This takes fifteen seconds and it is the single highest-value habit in this module.

A useful ritual before closing a long thread:

Summarise, in under 150 words, the conclusions we reached and the constraints I set. I am going to paste this into a fresh conversation.

Then check the summary is right, and paste it into a new window. You have kept the signal and dropped the accumulated noise.

Restate constraints periodically

In a working session, every ten messages or so:

Reminder: 1,500 words, undergraduate history, argument is that the reforms were driven by fiscal pressure rather than ideology, British spelling, no new sources.

Cheap insurance against truncation and drift, and it forces you to keep the constraints explicit in your own head.

One thread, one job

Keep separate conversations for separate jobs, in the way you would keep separate documents:

• One for the essay you are writing.
• One for the problem set.
• One for the reading you are trying to understand.
• One scratch thread you throw away daily.

Mixing them costs you quality because everything conditions on everything else. Separating them also makes your process record legible later, which matters when you write a disclosure statement.

Custom instructions and saved prompts

Most tools let you set standing instructions that apply to every new conversation. This is where the boilerplate belongs, so you stop retyping it:

I am a second-year undergraduate in economics. British spelling. Be direct; no preamble, no summaries of my question, no offers of further help. When I ask about a specific figure, source or paper, say plainly if you are not confident rather than producing one. When I ask you to check my reasoning, look for the weakest step rather than confirming.

Two warnings. First, standing instructions are also standing context: a note that you are a second-year will simplify things you later need in full, so revisit it. Second, whatever you write there is stored on the provider’s servers along with any memory feature, so keep it professional and impersonal.

Beyond that, keep a plain text file of prompts that worked: the quiz-me prompt, the marker prompt, the explain-back prompt. A file of six good prompts you actually reuse beats any purchased prompt library.

The signal that a thread is finished

When you find yourself explaining to the model something it already agreed to twice, the thread is over. Do not argue with it. It is not being stubborn; it is being conditioned on a context that now contains the argument. Close it, start again, and give the fresh window a clean statement of where you got to.