Tell when a conversation or document is too long to get good answers, and know what to do about it.
When an AI chat stops following rules you gave it near the start, or misses something you know is in a long file, start a new chat and open it with a short note of the goal, the rules and where the work stands. For long files, give the assistant smaller pieces and ask it to show where its answer came from.
Say you spend an evening revising your CV with an assistant. In your opening message you set two rules: the CV must fit on one page, and the two years you spent at a start-up that folded should not appear. The assistant sticks to both for the first hour. Forty messages later, the draft runs to a page and a half, and the start-up is back in with a glowing description.
You did not change your mind, and the assistant has not chosen to ignore you. Your opening message has, in practice, faded out of what the model is working from.
Lesson 5.1 described the context window as everything the model can see at once. Each message you type and each reply you get takes up room in it. Pasting in documents and getting long, detailed answers fills that room faster than most people expect.
Once a conversation is longer than the window, the app has to deal with the overflow, and different products handle it differently. One approach drops the oldest messages, so the model no longer sees them. A second replaces older parts with a short summary the model writes itself, which keeps the gist and loses the detail. A third tells you the chat is full and asks you to start a new one. The first two are the risky ones, because they happen silently. You can still scroll up and see your early messages on screen, but the model may no longer see them.
Early instructions are the ones most likely to be dropped, or compressed into a summary that loses a qualifier. In the CV chat, the summary could keep "one page" and lose "and not the start-up".
A second problem applies even when everything fits in the window. Models do not weight every part of a long text equally. In practice they handle the start and end of long material better than the middle, so a detail buried halfway through a report or transcript is the most likely to be missed.
An operations executive at a Jurong logistics firm uploads a fifty-page vendor contract and asks whether there is a penalty for late delivery. The clause is on page 27. The assistant summarises the payment terms from the opening pages and the termination terms from the end, then either says nothing about penalties or says there are none. The clause was in the window, and the model did not use it. A large window helps you get long material in, but it does not guarantee that every line is read carefully. Lesson 5.4 has you test this on a real document.
There is rarely a warning that a chat has drifted, meaning an earlier instruction or correction is no longer being followed. You have to spot it from the answers. There are three common symptoms:
An early constraint stops being followed: a word limit, a tone, something to leave out, or the format you asked for. A mistake you already corrected comes back, such as the wrong client name or a figure you fixed an hour ago. Answers about a file ignore part of it. Questions about the opening pages get good replies, while questions about a middle section get vague or wrong ones.
Any one of these is a signal to stop pushing the same chat further. Repeating the instruction may help for a message or two, but in a long chat it tends to slip again. A rough rule: when a chat feels like it is arguing with you, the part where the rules were agreed has often gone.
Open a new chat and give it a carry-over note: a short summary in three parts, made up of the goal, the rules that must hold, and the current state of the work. You can write it yourself, or ask the old chat to draft it and then check it before you use it, because the old chat's summary may already have lost your early rules. Put the note at the top of the new chat, where the essentials stay in plain view and get attention.
For the CV, four lines are enough: you are applying for a regional marketing manager role, keep it to one page, leave out the 2021 to 2023 start-up role entirely, and the latest draft pasted below.
Long documents need a different approach. Give the assistant only the section that matters, or work through the document a part at a time and ask the same question of each part. When the answer has to draw on the whole document, ask the assistant to name the page or quote the passage it used, then check that page yourself to confirm it read the relevant part.
Most people have at least one old chat like that sitting in their history. Go and find yours, because looking at a real one shows you where your own instructions started to slip.
Pick a long chat from your history. Scroll up to an early instruction and check whether the latest replies still follow it. Then write down what you would carry over into a fresh chat.
Junxiong-WFG Organisation is an authorised representative of AIA Financial Advisers Private Limited (Reg. No. 201715016G).