Level 2: Context Engineering - How AI Becomes a Co-Worker
Published April 28, 2026 · The Human Question · By Rob Gonzales, CPA, PhD
Once you have learned to write a clean prompt, the next thing you discover is that prompts alone are not enough.
You can write the most beautiful, structured prompt in the world. Role, objective, context, task, audit. All five parts in place. And the answer can still be wrong, because the AI is missing something it had no way to know.
That something is context.
The shift the field made in 2025
In 2024, almost every conversation about getting better AI output was about better prompts. “Write a better prompt” was the universal advice. By the middle of 2025, the conversation had moved. The interesting practitioners were no longer talking about prompts. They were talking about context.
The reason is simple. Models got bigger. Context windows expanded from a few thousand words to hundreds of thousands. Frontier systems can now hold an entire book or a full client engagement file in a single conversation. The bottleneck is no longer how clever your sentence is. It is what the model can see.
Prompting is what you say. Context is what the AI knows when you say it.
What context actually is
Context is everything inside the AI’s working memory at the moment you ask the question. That includes:
1. Files you uploaded — a 10-K, a contract, a textbook chapter, a prior-year memo.
2. Notes you pasted — your engagement letter, your style guide, your team’s standards.
3. The conversation so far — every question you have already asked and every answer it has given.
4. Project files in tools like Claude Projects or ChatGPT Custom GPTs — persistent context that survives between conversations.
5. Memory features — the new layer where AI remembers facts about you across sessions.
Each of these is a layer. The more thoughtfully you stack them, the smarter the AI looks. The layers are doing the work.
The error model, revisited
In Lesson 1, I described prompting as an error-reduction problem. Context engineering is the same problem, scaled up.
Output quality = context quality × prompt quality × model capability.
If any of those three terms is near zero, the whole thing collapses. A genius prompt with no context produces a confident, plausible, wrong answer. Plenty of context with a vague prompt produces a meandering essay. The art is balancing all three.
How real teams build context
I recently had a few clients ask me how their teams could go from “sometimes useful” AI to “daily-driver” AI. The answer was not better prompts. It was a better context system.
Here is what we built together, in roughly the order it should be built.
1. A master context file
One document, kept in the project folder, that any new AI session can be opened with. It says who the team is, what the work is, what the rules are, what the deliverables look like, and where to find the source materials. Drop it in once at the start of a session. The AI reads it and now it knows everything that was previously trapped in someone’s head.
2. A folder structure the AI can read
If your project lives across forty different OneDrive folders with cryptic names, the AI cannot help you. Build a numbered, predictable structure: source materials, reference standards, work in progress, deliverables, archive. Same labels every time. The AI learns the layout once and then just works.
3. A short list of canonical examples
Three to five examples of what “good” looks like for whatever you are producing. A finished question. An approved memo. A polished email. Drop them into the conversation when you brief the AI. It will pattern-match better than any instruction you could write.
4. A brand-voice or style note
One paragraph, sometimes one page, on how your team writes. Avoid these words. Prefer these. Two-space sentences. No em dashes. Whatever your house style is. The AI will respect it as long as it can see it.
5. A way to keep the context fresh
Context decays. A master file written in January is wrong by April if the project has moved. Version it. Save a v2, then a v3. Never overwrite. Treat the context file like the most important document in the project, because for an AI workflow, it is.
The shift you can feel
The first time you do this well, something changes. You stop typing long prompts. You stop re-explaining your project at the start of every conversation. You drop a context file, you ask one short question, and the AI responds like a colleague who has been on the team for six months.
That is when AI stops feeling like a chatbot and starts feeling like a co-worker.
Three rules to take with you
1. Curate, do not dump. More context is not always better. The AI reads what you give it. Give it the right things, not all the things.
2. Make context durable. If you have to retype it every session, it is not context. It is just typing. Save it as a file the AI can re-read.
3. Treat context like an asset. The best context files I have built are the most valuable artifacts on those projects. They get versioned, reviewed, and improved like code.
The real lesson
Prompts are sentences. Context is the world those sentences live in.
If you want AI to stop sounding generic, stop writing better prompts. Build a better world for it to read.
Next up: Level 3 — Workflow Engineering: When AI Becomes a Process