AlvoreachStart free

Free resource · Cheat sheet

Make it fit the context window

AI is not forgetful — it has a fixed working memory, and long jobs quietly overflow it. Here is how to spot the overflow, work in parts, and hand work between chats without losing the plot.

Free, and free of a catch. See the whole library.

The window, in one picture

Everything in a conversation — your messages, the AI's answers, every document you paste — shares one fixed working memory, called the context window. While there is room, the model can use all of it. When the conversation outgrows the window, the oldest material stops being visible — and nothing announces it. The model does not say 'I have lost the start'; it just carries on, minus the start.

what the model can still seefallen out — silentlystart of the conversationnow
Nothing warns you when the oldest part falls out — the conversation just carries on without it.

The three signs you have overflowed it

You cannot see the window, but you can see the symptoms. Any one of these in a long conversation means the start has probably gone.

  1. 1It asks for something you already gave it — a detail, a file, a decision from earlier.
  2. 2Instructions from the start stop being followed — the tone, format or limits you set drift back to defaults.
  3. 3It gets confidently vague about the early material — summarising the first part of a document in a way that sounds right but does not match what you pasted.

Work in parts, carry a summary

For anything longer than the model can comfortably hold — a big report, a contract, a day of notes — do not paste it all and hope. Feed it in parts, and make the model maintain a running summary that travels with you. The summary is small enough to stay in the window even when the parts themselves have fallen out.

Start of a long job
I am going to work through a long supplier contract in parts. For each part I paste, do two things: first, pull out anything about our obligations, deadlines, penalties and renewal terms; second, update a running summary of everything so far, in under 150 words, and put it at the end of your answer. I will paste that summary back to you with each new part.
The running summary is the trick — it is the document's memory, carried by hand. Swap the contract and the things to pull out for your own document and goal, then paste part one straight after.

Re-anchor before the answer that matters

Instructions age badly in a long chat: the rules you set at the start are the first thing out of the window. So when the important ask comes — the final draft, the recommendation — restate the brief in the same message. One line is enough. It costs you ten seconds and saves the 'why has it forgotten the format' round-trip.

Paste above your final ask
Before you answer, the brief from the start still applies: we are writing a one-page funding proposal for the town council, in formal but plain English, with no jargon and no promises about timelines. With that in place: give me the final draft.
Swap the middle for your own brief — job, format, limits — and your real request for the last line. It reads as fussy; it works because the newest message is the one thing guaranteed to be in the window.

Know when to start fresh

Past a certain length, a conversation is carrying more history than help — old drafts, dead ends, corrections to things that no longer matter. When the job changes shape, a clean chat with a tight hand-over beats soldiering on. Ask the old chat to write the hand-over for you.

The hand-over
Summarise this conversation as a hand-over note for a fresh start: the goal, the decisions we have made, the current state of the work, and what is left to do. Under 200 words, no history — just what a newcomer needs to carry on. I will paste it into a new chat.
The new chat starts with a full window and only the material that still matters. This one prompt is the difference between a long job degrading and a long job finishing.

The simplest rule of all

If you would have to scroll to find it, assume the model may have lost it. Either paste it again, fold it into the running summary, or start fresh with a hand-over. Those three moves cover every long job — and none of them needs you to know a single number about tokens.

More like this, as we write it.

The library is still filling. Start free and we will tell you when the next resource ships — one email, no spam.

Start free