From Chats to Systems
Configure the Machine
Read-aloud narrates the unit in a natural voice, paragraph by paragraph. It needs an account.
In short
The assistants already remember you, by accident. The operator's first move is replacing accidental memory with deliberate architecture: standing instructions, grounding files, and scoped memory.
The Hook
Open your main AI assistant right now and type: "What do you remember about me?"
You'll get a list. Some of it accurate. Some of it stale. A preference you mentioned once, half joking. A project you abandoned in March. Your name, probably. Nothing about the documents that actually define your work, and no wall between your chemistry homework and your job applications. Nobody chose that list. It accumulated. (And if you got nothing back at all, your account is new or memory is off. Keep that blank page in mind; it's the other starting point this unit builds from.)
The old complaint about AI was amnesia: it forgot you between chats, and you re-explained yourself every morning. That problem is gone. The major assistants now remember you between conversations, most of them by default. But look at what they remember. It's a junk drawer.
So the gap has moved. It's no longer remembered versus forgotten. It's remembered by accident versus briefed on purpose. Closing that gap is the first skill of operating AI instead of just using it, and it's where this course begins.
The whole course fits on one diagram. Each unit climbs one rung, from writing a single good prompt to running systems that work while you sleep.
Before any theory, touch the top rung. This is a working miniature of the kind of system you'll be able to build by Unit 08: a small social media pipeline that researches, drafts, and waits for a human at exactly one moment. Press run, and hold the gate yourself.
Marra Coffee runs a small social studio. Five steps, one of them yours. Everything here is simulated.
Four steps ran on their own. One needed you. Hold onto that ratio; the whole course is about earning it safely. Today you build what that system stands on: the workspace.
Here's the mental model for this unit: build the machine that writes the brief.
A chat is a briefing you deliver by hand, every time, from scratch. A A configured, persistent home for one area of your work: standing instructions, attached files, and its own memory. Every chat inside it starts pre-briefed. is a machine that delivers the briefing for you. The instructions are already standing. The right files are already attached. The relevant history is already scoped. You walk in and start at the interesting part.
Nothing about the engine changed. At question time, the model does what it always does: The way AI generates all of its text: it looks at everything so far, guesses the most likely next chunk, and repeats one piece at a time. Uploads and memories improve the input; they don't change the engine.. What changed is what stands in front of the engine. Since the engine is fixed, the input is the only variable you control, and a workspace is that variable, engineered instead of improvised.
Think of the difference between a freelancer and a hire. A freelancer you brief from zero every single morning: who you are, what the project is, what good looks like. A hire has an office. Standing instructions taped to the monitor, the right folders in the drawer, a memory of last month's decisions. Same brain in both cases. The hire produces better work by Tuesday because the briefing became furniture. A workspace is you giving the AI an office.
Every workspace is built from three layers, and each layer has an accidental twin you're replacing:
Standing A short block of text that applies to every chat inside a workspace: who you are, the voice you want, what to always do, what to never do. Written once, applied forever.. The deliberate version: a written block defining role, voice, and rules. The accidental version: whatever tone your last few messages happened to set.
Restricting the AI to answer from sources you provide (your files, your notes) instead of its general training. Grounded answers cite where they came from, which makes them checkable. files. The deliberate version: the documents that define your work, attached as the source of truth. The accidental version: whatever the model half remembers from its training data about topics like yours.
Scoped What an assistant carries about you between conversations. Scoped memory belongs to one workspace and stays there. Unscoped memory is the account-wide junk drawer from the hook.. The deliberate version: each workspace keeps its own history, walled off from the others. The accidental version: the junk drawer, where your job search leaks into your homework help.
Here's the landscape as of August 2026. Claude Projects (claude.ai) hold instructions, files, and a separate memory space per project; free accounts can create up to five. ChatGPT Projects (chatgpt.com) work the same way on every plan, with project instructions that override your global settings and an option, chosen at creation, to keep memory project-only; free accounts attach up to five files per project. Gemini Gems (gemini.google.com) are custom assistants with standing instructions, inside an app that passed 900 million monthly users by mid 2026. NotebookLM (notebooklm.google.com), which Google folded into its AI plans and renamed Gemini Notebook in 2026, is the purest version of the idea: a workspace that answers only from sources you upload. Its free tier runs to roughly 100 notebooks with 50 sources each, and a single source can run to hundreds of thousands of words.
Every tool in that paragraph has a real free tier. This entire unit runs on zero budget.
NotebookLM went viral in late 2024 for a party trick: turning any pile of documents into a two-host podcast. But the trick was never the point. The point was the architecture underneath: an assistant that answers only from the sources you give it, and pins a citation to the exact passage behind every claim. Ask it something your sources don't cover and it's built to say so instead of improvising.
By 2026 the product had grown into the workspace layer of Google's AI lineup, renamed Gemini Notebook, with students running entire courses through it: lecture notes in, grounded study answers out, each one checkable against the page it came from.
The lesson for an operator: grounding is a powerful control, and it doesn't retire the oldest rule of working with AI, which is to verify anything that matters. You've just moved from checking what the AI invented to checking what it did with your material. A citation is an invitation to look, not permission to skip looking.
A workspace narrows what the AI draws on. It does not guarantee what comes out. Three things still fail: retrieval can miss the passage that mattered, synthesis can bend what a source actually said, and memory goes stale the moment your situation changes. The junk drawer doesn't disappear when you curate it; it just becomes a drawer worth auditing.
So verification doesn't retire here. It moves upstream: from fact-checking finished answers to auditing what your machine feeds in. You'll build that audit into a habit in the Operator Moves below.
Free path: every step below runs on free tiers: Claude (up to five projects), ChatGPT Projects, Gemini Gems, and NotebookLM.
Step 1, the audit. Open the assistant you use most and run:
PromptWhat do you remember about me? List every item, and for each one, tell me where it came from if you can.
Grade the list: accurate, stale, or missing. If the list is empty or memory is off, that's the other half of the same lesson: nothing persists unless something is built. Either way, you now know your starting point.
Step 2, watch one workspace get built. Through this course, one worked example runs in every demo: the Studio, an automated social account for a fictional neighborhood coffee shop called Marra Coffee. Here is its entire standing instruction block:
PromptYou are the content assistant for Marra Coffee, a small neighborhood coffee shop. Voice: warm, short sentences, no emojis. Content pillars: new drinks, behind the counter, neighborhood stories. Never use: "game changer," "elevate," "coffee lovers rejoice." Always: one concrete detail per post, and end with a question to the neighborhood. The attached files are the source of truth; prefer them over general knowledge.
Create a project in any of the three assistants, name it "Marra Coffee Studio," paste the block into its instructions, and ask it for one post about a new cardamom latte. Notice what you didn't have to say. (Delete the demo project afterward if you need the slot; Claude's free plan allows five.)
Step 3, now yours. The Studio is the demonstration; your life is the assignment. Pick one domain you actually work in (a course, a job, a hobby project) and fill this skeleton in your own words:
PromptYou are my assistant for [domain]. Who I am: [one line] Voice: [three adjectives] Always: [two rules] Never: [two rules] The attached files are the source of truth; prefer them over general knowledge.
Create the workspace, paste your block, attach one or two real files.
Step 4, ground it. Open NotebookLM, create a notebook for the same domain, and upload two or three real sources. Ask a question that only those sources can answer. Then click the citation and check the passage: does it actually say what the answer claims? That check is the new shape of verification.
Step 5, the head-to-head. Take one real task from that domain. Run it once in a bare chat with no setup, and once inside your new workspace. Put the outputs side by side.
Operator Moves
The monthly memory audit. Once a month, ask each assistant what it remembers about you. Delete what's stale. Anything durable and true gets promoted out of memory and into a workspace's standing instructions, where you control the wording. Memory is where facts about you happen; instructions are where they're decided.
The portable instruction block. Keep each workspace's instruction block in a plain note outside any AI tool, under 150 words. Tools change, accounts reset, better assistants launch. With the block in your pocket, you can hire any AI on earth into your context in sixty seconds.
Scope by domain, never by tool. One workspace per area of your life, and no "everything" workspace, ever. The wall between your workspaces is a feature you built, and it's exactly what the junk drawer never had.
Everything else in this course runs on what you just built. In the coming units you'll hand work to research agents, terminal agents, and eventually automations that fire while you sleep. Every one of them inherits its context from workspaces like these. A sharp workspace makes each of them sharper. A sloppy one turns them into machines for producing confident garbage at scale.
There's also a quieter payoff. Re-explaining context isn't just tedious; it's a cost you pay again on every task, in time and in tokens. The machine that writes the brief pays that cost once.
And it changes your relationship with the tools. The junk drawer happens to you. The workspace is something you decided. That difference, deciding what the AI knows instead of discovering it, is the difference this whole course is named after.
Three Workspaces
Build the ground floor of your operator toolkit. Later units will use these directly.
- Pick three domains from your real life (a course, a job or job search, a project, a hobby with output). Not hypotheticals: areas where you'll actually do work this month.
- Write three instruction blocks using the skeleton from the demo: role, who you are, voice, two always rules, two never rules. Under 150 words each. Save all three in a plain note outside any AI tool.
- Build the workspaces in the tool of your choice (Claude Projects, ChatGPT Projects, or Gemini Gems; all free). Paste each block, and attach at least two real files to at least one workspace.
- Make one memory decision per workspace: should this workspace share account-wide memory or stay scoped? Write one sentence justifying each choice.
- Run the head-to-head in each workspace: one real task, bare chat versus workspace. Note one concrete difference per pair, in one sentence each.
- 1The assistants already remember you, by accident. The operator's first move is replacing accidental memory with deliberate architecture: standing instructions, grounding files, and scoped memory.
- 2The engine is unchanged next-token prediction. Workspaces win by industrializing the input, which beats any clever phrasing on any single prompt.
- 3Grounding narrows the AI to your sources and makes checking fast; it doesn't retire checking. You now verify what the AI did with your material.
- 4One workspace per domain, blocks saved portably, memory audited monthly. These three workspaces are the ground floor every later unit builds on.
Type: Video Title: How to Study Anything with NotebookLM (Full System) URL: https://youtube.com/watch?v=TBKjg9ObgY4 Description: A complete grounded study system built end to end: sources in, briefed workspace out. Watch someone live an entire workflow inside the architecture you just built, then steal the parts that fit your life.
| Type | Title | URL | Description |
|---|---|---|---|
| Docs | Claude Help Center, "How can I create and manage projects?" | support.claude.com/en/articles/9519177… | Projects on all plans; free accounts can create five |
| Docs | Claude Help Center, "Use Claude's chat search and memory" | support.claude.com/en/articles/1181727… | Memory on free, Pro, and Max, with a separate memory space per project |
| Docs | OpenAI Help Center, "Projects in ChatGPT" | help.openai.com/en/articles/1016952… | Project instructions, files, and project-only memory explained |
| Reference | OpenAI announcement: Projects for Free users (Sept 2025) | x.com/OpenAI/status/19633… | The rollout that made workspaces free for everyone |
| Article | TechCrunch, "Google's Gemini app has surpassed 750M monthly active users" (Feb 2026) | techcrunch.com/2026/02/04/googles-… | The scale of the assistant carrying Gems; 900M+ was disclosed at I/O in May 2026 |
| Tool | NotebookLM (Gemini Notebook) | notebooklm.google.com | The grounded workspace; free tier includes 100 notebooks with 50 sources each |
| Video | "The Ultimate Google NotebookLM Guide (2026 Full Tutorial)" | youtube.com/watch | Beginner-to-pro walkthrough of the grounded workspace, January 2026 |
| Video | "How to Study Anything with NotebookLM (Full System)" | youtube.com/watch | A full grounded workflow, lived end to end |
| Article | Anthropic, "Effective Context Engineering for AI Agents" | anthropic.com/engineering/effecti… | The design discipline behind everything this unit builds |
Last updated: August 9, 2026. Tool names, tiers, and limits were verified on this date. The mental model outlasts all of them.
Marking a unit complete, the Prove It check and your place in the course all need an account. The reading stays free.