Context is the whole game: the six layers, your own layer, memory, context rot and projects

Humans hold only a handful of things in active working memory: classic estimates say about seven, newer ones closer to four. Modern AI models can hold hundreds of thousands of words at once, sometimes a million. About 750,000 words is the first four to five Harry Potter books, or several days of continuous speech. The model can read all of it before answering. But it can only read what you give it. ¹

The six things on the desk

Context is everything that ends up in the model's window for a given response: the system prompt the product set, the descriptions of any tools it can call, your prompt, the chat history of this conversation, any files you uploaded, and, newest of the six, a short profile the tool has written about you. Whatever is on that desk right now is all the model knows, and anything not on the desk does not exist for this answer. ¹

The window is not empty when you arrive

When you open a fresh chat, you might think you are starting with a blank surface. You are not. Before you type a single character, the company that built the tool has already placed a set of instructions on the desk. You will never see them in the chat, but the model reads them before it reads anything you write. Think of a restaurant owner briefing a new waiter before the first customer sits down: "Be friendly. Recommend the daily special. If someone asks about allergens, always check with the kitchen, never guess." Engineers call those invisible instructions the system prompt. ¹

What is typically in that briefing: how to behave, what to refuse, what tone to use, when to add disclaimers, what tools it can call. This is why Claude, ChatGPT and Gemini feel different even when you ask the exact same thing. The personality you sense is not baked into the model; it is baked into the instructions the company loaded. Now you know why the AI is polite even when you are rude, refuses certain requests, adds disclaimers you never asked for, and sometimes apologizes more than seems necessary. These are not personality traits. They are instructions. ¹

Your own layer

The company's system prompt is fixed, but most tools let you write your own instructions that load alongside it into every chat. When you write "I am a nurse, assume clinical vocabulary" in your tool's instruction settings, you are writing your own line in the system prompt. The model reads it before every response, which is why it sticks without you repeating it. ¹

A teacher sets her instruction to: "I teach Grade 5 science. Explain everything at a 10-year-old's reading level. Never use jargon without defining it first." Every new chat now has both briefings on the desk, the company's and hers, before she types a word. She never has to say "I'm a teacher" again. ¹

Keep your own layer short, and prune it

Because this layer loads before every chat, it is tempting to keep adding to it: one more line each time the AI does something you did not want. A year later you have twenty lines, and some quietly contradict each other. Anthropic ran into this inside its own products and in July 2026 deleted most of the standing instructions it had accumulated, with no loss in quality: modern models read the situation well enough that many old rules were only getting in the way. ^{1 2}

Two habits keep this layer useful. Write what the AI cannot work out on its own, your job, your audience, your hard constraints, and not what it would have done anyway. And read the whole thing every few months, asking of each line: if I delete this, will the AI actually get something wrong? If not, delete it. The same test applies to project instructions, and in this record it applies to every agent's instructions and constitution. ¹

Same model, same question, different answer

• Bare prompt: "pros and cons of studying physics versus zoology." You get generic high-school-counselor advice.
• Context-rich prompt: the same question, plus your career assessment results as a PDF and a screenshot of your high-school schedule. Now the AI can talk about your specific aptitude profile and course history.

The difference is the context, not the cleverness of the prompt. The discipline: before you press send, ask yourself what a smart new colleague would need in front of them to answer this well. Then attach those things. The colleague will read everything carefully; they will not guess what you did not tell them, will not search your filing cabinet, will not infer your industry, your team's history or yesterday's email thread. ¹

The book's example: a 7th-grade teacher asked for "a lesson plan on the water cycle" and got a generic textbook plan. The next day she attached her syllabus (what came before and after), last week's graded worksheets (which concepts had landed and which had not), and the school's standardized test format. The new plan opened with a review of the two weak concepts, threaded the material through the May test format, and closed with a check matched to the syllabus's next topic. Same model, same teacher, same subject. ¹

The checklist before any non-trivial prompt

Five lines of context, properly chosen, beats five paragraphs of cleverness. The same territory can be sliced as a five-slot template, role, context, task, constraints, output format; the checklist sorts by what you might forget to attach, the template by where each line goes. The habit underneath is identical. ¹

The sixth layer: the AI now writes notes about you

All three tools now write a short summary of who you are and how you like to work, and load it at the start of every new chat. This does not contradict the stateless fact, and the reason matters: memory does not give the model a memory. It is a note the tool keeps about you and places on the desk. A sixth layer on the stack, not an exception to it. ¹

Menu names move around, and Gemini's version needs a personal Google account with activity history on. Check your own settings rather than trusting the table. Three habits and one warning: ¹

• Read what it has stored, once. Some of it is sharper than expected, and some is a preference you mentioned once in passing that has been shaping every answer since. You can delete any line.
• Correct it out loud. "You are assuming I still work in retail; I do not." That updates the note without opening settings.
• Use the clean-slate mode when the context would only mislead: a one-off question far from your usual work, or a topic you do not want shaping future answers.
• The warning, for anyone with a duty of confidentiality. A memory note is a place client and student details can quietly accumulate and resurface in an unrelated chat months later. Keep identifiable details out of memory; do that work in a clean-slate chat or a scoped project.

If the AI keeps making the same wrong assumption in every new chat, stop rewriting your prompt and go fix the memory. ¹

Context rot

Context windows are large, but not infinite, and recall degrades inside them. The biggest practical mistake: one very long conversation across many unrelated topics. The workout context is still in there, distracting the model, when you ask it to debug a spreadsheet. Rule of thumb: when the topic changes, start a new conversation. ¹

Symptoms that a conversation has gone stale: the AI references earlier parts of the chat that have nothing to do with what you just asked; its answers get longer and vaguer, with more hedging; it contradicts a constraint you stated five turns ago; it apologizes repeatedly without making progress. ¹

What is happening has a name: once a conversation gets long enough, most tools quietly compact the older parts, summarizing the early turns into a short paragraph and replacing the originals. The narrative survives; the specifics do not. The library you named three hours ago, the naming convention you agreed on, the constraint in turn four, any of these can disappear into the summary. So: a chat window is working memory, not storage. Anything that needs to survive one long session belongs in a project, an attached file, or a note you can re-paste. ¹

When you see the symptoms, resist one more clarifying prompt; that adds tangled context to a tangle. Start the new chat, paste in the one or two facts that matter, and continue. The reset is almost always faster than the rescue. If the dead chat produced something worth keeping, save it to a file first. ¹

Projects: front-load context once instead of every time

If AI needs to be briefed like a colleague every time, that is a lot of repeated typing. The answer most tools ship is projects: a workspace you set up once, with the files, instructions and audience that always apply to a kind of work, so every chat inside it inherits that setup. When to make one: the moment you notice you have pasted the same files, the same audience description or the same constraints into two or more chats on the same topic. ¹

• A tax filing project with last year's return and the forms, and "Assume I am a US filer with one dependent. Always show your math."
• A kids' school project with the syllabus and calendar, and "Always check the date against the calendar before answering."
• A writing voice project with three samples of your writing, and "Match the cadence and word choice of the samples. Do not add hedging or qualifiers I did not use."

Inside a project, "start a new chat" no longer loses what the AI knows about your situation; it loses only the noise of the previous conversation. You reset the chat, not the context. ¹

Three tools, one idea: Claude calls it Projects, ChatGPT calls it Projects, Gemini calls it Notebooks, which sync with NotebookLM. Claude and ChatGPT Projects tilt toward instructions and behaviour: you set the voice, the role, the rules, the audience, and the model holds that persona across every chat. Gemini Notebooks and NotebookLM go further on the source side: PDFs, Docs, URLs, videos, audio, every answer grounded in those sources with citations, and the workspace flows both ways, so last week's chat is one more source this week's chat can cite. Rule of thumb: reach for Notebooks when the workspace will grow over time; reach for Claude or ChatGPT Projects when it is built around a persona or a set of instructions. As of mid-2026 all three have a free tier; Claude caps how many projects you can have (up to 5), ChatGPT caps how many files each project holds (5 on the free plan, 25 or 40 on paid plans), and Google's are free with paid tiers raising source limits. Plan your structure around whichever cap will bite first. ¹

Continue to reasoning and sycophancy.

Footnotes

1. AI Prompting 2026: 13 Concepts, 80% of Real Use, Part 2, Concept 4 "Context is the whole game", including the dropdowns on pruning, context rot and projects. The AI Agent Factory, Panaversity, 2026. ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸ ↩⁹ ↩¹⁰ ↩¹¹ ↩¹² ↩¹³ ↩¹⁴ ↩¹⁵ ↩¹⁶ ↩¹⁷ ↩¹⁸ ↩¹⁹ ↩²⁰ ↩²¹

2. The new rules of context engineering for Claude 5 generation models, Anthropic, July 2026, as cited by the book. ↩