AI prompting guide 2026

Most people use AI like a search box. They type a short question, skim the answer, and move on. That works for trivia; it fails for everything that matters. Power users do something different: they brief AI the way they would brief a smart-but-new colleague, with files, context, constraints and a clear ask. They expect three options instead of one. They argue. They iterate. They check the work. The gap between a novice prompt and a power-user prompt is not cleverness; it is a handful of habits anyone can learn in an afternoon. ¹

This room is that afternoon, kept in full: thirteen concepts in four parts, from the book, with its tables, examples and prompt templates. It is here for two readers. A person, who wants better answers from any chat tool. And an agent working from this record, which must follow these rules whenever it writes an instruction, a system prompt, a tool description, a spec or a brief for a coding agent.

The one fact, and the one insight

One fact underlies everything: the model is stateless. It has no memory of its own between turns, and answers each time using only what is in its context window right now. Everything else is downstream of that. So one insight runs through every concept: almost every "advanced technique" is one of two moves, getting the right context in, or keeping the wrong context out. The model sees only what is in its window for this response. Your job is to control what goes in. ¹

What changed since you last looked

If you used ChatGPT in 2022 or 2023 and decided it was a clever toy, the tool you remember is not the tool you have now: ¹

• Context windows grew by roughly 1000x. A 2022 model held a few thousand words; a 2026 model holds hundreds of thousands, sometimes a million. A whole book, several days of speech, a folder of contracts fit in a prompt.
• Reasoning became real. "Think step by step" used to be a magic phrase. Now models have explicit thinking modes that run for seconds, sometimes minutes. A year ago the hardest task AI could reliably finish took a person a few minutes; today it is something that would have taken an hour or more.
• Web search became a built-in tool. The model decides when a question needs fresh information, searches, reads a few pages, and uses what it found.
• Code execution became a built-in tool too. The model can write a small program, run it, and use the result. Both tools are mostly invisible, so most users cannot tell whether an answer came from memory, a web page, or a calculation. Once you notice, your prompts get sharper: "did you actually search for this?", "run the numbers, don't estimate."
• Multimodal stopped being a sidebar. A photo, a PDF, a spreadsheet, a voice memo, a folder of files, all in one stream.
• The tools started remembering you. They write a short profile of who you are and load it before every chat. The model did not gain a memory; the product put a note on the desk.
• Desktop apps appeared that find your files, draft emails and update spreadsheets with permission. That is delegation, not chat.
• Command-line agents appeared for developers, living in the terminal, reading a whole codebase, editing many files, running tests.

If your mental model is out of date by even eighteen months, you are using these tools at maybe 20% of what they can do. ¹

How this record applies the guide

Whenever an instruction is written from this record, for a person or for an agent, these are the rules it follows, each from the page named:

1. Brief like a smart new colleague who does not know you yet: context, constraints, a clear ask. How AI knows things.
2. Write what the model cannot work out on its own, and nothing it would have done anyway; read standing instructions every few months and delete any line whose removal would not cause a mistake. Context is the whole game.
3. Attach the document, state the constraint, name the audience and the format. Five lines of context beat five paragraphs of cleverness.
4. Name the sources you trust, ask for citations inline, ask it to flag what it could not verify. How AI knows things.
5. Ask for thinking on hard, multi-input questions, and not on quick ones. Reasoning and sycophancy.
6. Frame neutrally. Replace find, defend, confirm, prove with evaluate, compare, critique, find any. Use a rubric; force a number per criterion, and ask for the change that would raise it.
7. Run the loop: load context, ask for three to five options, give explicit feedback, ask again, expand only the winner; outline before drafting; research before either. The brainstorm-iterate loop.
8. Match the grip to the kind of ask: loose when you want options, tight when you want execution; never specify the conclusion of research or analysis.
9. When numbers are involved, make it run code and show the code. Beyond text.
10. Scope permissions to the smallest folder, and grow them with track record. Safety, cost and checking.
11. For work where being wrong is expensive, have a second model from a different family grade it against the same rubric.

The Build your agent room leans on this directly: an agent's instructions, its constitution, its spec and every prompt pasted to a coding agent are prompts, and they are written under these rules.

The pages

The thirteen concepts in one line each

• 1. The gap between a novice and a power user is a handful of habits: brief AI like a smart new colleague, with context, constraints and a clear ask.
• 2. AI knows things from a snapshot of the internet; it learned by reading text about the world, not by experiencing it, so it is strong on common topics and weak on obscure or recent ones.
• 3. Three retrieval modes: pretrained, web search, deep research. Your wording steers which one fires.
• 4. The model has no memory of its own; the context window is its working memory for this response. What you put in that window is the single biggest determinant of answer quality. Projects front-load it once; the tools' own memory adds a sixth layer you did not write, so read it and prune it.
• 5. Modern models can think hard for seconds or minutes if you ask.
• 6. Models are biased toward agreement. Neutral framing and rubrics neutralize most of it; forcing a 1 to 10 score per criterion, with the change that would raise each score, neutralizes the rest.
• 7. The iterate-with-explicit-feedback loop is the highest-leverage habit on the page. Grade each stage out of 10 and re-iterate until the score plateaus; the AI does not get to declare you finished. Match the grip to the kind of ask.
• 8 and 9. AI can see images, work with audio both ways, and build small apps; the running app is an artifact you can iterate on, share and embed.
• 10. AI can write code and run it on your data, but does not always do this automatically. Ask explicitly, and verify the code actually ran.
• 11. File-aware desktop apps are a new category. Scope permissions tightly until you have used them safely.
• 12. The right tool for a job changes every few months. Know the distinct families, expect new ones, and check a leaderboard monthly.
• 13. When no human expert is in the room, making models grade each other, across different families, is the closest thing to an objective quality signal. ¹

Underneath all of it is one move, repeated in a dozen disguises: get the right context in, keep the wrong context out. ¹

Begin with how AI knows things.

Footnotes

1. AI Prompting 2026: 13 Concepts, 80% of Real Use: the introduction, "A short note on what changed since you last looked", and "A short recap before you try the prompts". The AI Agent Factory, Panaversity, 2026. ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶