Setting up your AI partner the right way means stacking four layers underneath every conversation so the tool stops starting from zero. Layer 1 is custom instructions, the lobby every chat walks through (your role, your audience, your voice, your rules). Layer 2 is memory, the notebook your tool keeps on you (some of it auto-written, all of it editable). Layer 3 is projects, the conference rooms with their own custom instructions and their own files (great for client work, content workflows, or any repeated job). Layer 4 is files, the source material on the table (brand guides, SOPs, past case studies, the authoritative stuff you don't want the tool to make up). Get those four layers loaded and the same prompts you've been typing for months start producing dramatically different work.
That's the build we walked through in Session 3 of our free 8-part series, Building Your AI Content System. The full 35-minute replay is up top. Hit play, or keep reading for the recap.
Why Most Humans Are Using AI Like A Vending Machine
If you've opened ChatGPT this week and stared at a blinking cursor thinking "this feels like an uphill battle, this thing doesn't really know me, I'm going to have to edit this output for an hour to make it usable," you're not alone and you're not broken. You're just running the default playbook most humans run.
You walk up to the machine. You punch in a prompt. Out comes one snack. You eat the snack, you post the snack on LinkedIn, you walk away. Tomorrow, same machine, same kind of snack, same walk-away.
That's not a tool problem. That's a system problem.
And the system is broken on purpose, because nobody ever told you that you could set it up. The good news is you can. Today. Without a computer science degree, without three hours of YouTube tutorials, without becoming a prompt-engineering nerd.
Here's the phrase George keeps saying when he coaches humans on this: garbage in, garbage out. And by the way, nothing in also equals garbage out. If what you put in is a short prompt with no context (or no context at all), the tool fills the blanks with the average of everything it knows. That's the internet's voice talking to the internet's audience. No wonder so many LinkedIn posts sound identical right now.
Context before complexity.
Get obsessed with the technique, not the tools. If you give the tool the context first, the complex stuff lands clean and the simple stuff feels easy. Skip the context and even the simple stuff feels overwhelming. That principle applies to teaching humans HubSpot. It applies to teaching humans ChatGPT. And it applies to teaching ChatGPT about you.
The Mental Model: Four Layers Stacked Like A Building
Before we touch a single setting, get this model into your head. Every reply your AI tool gives you is built on a stack of layers. Think of it like a building.
Layer 1 is custom instructions. That's the lobby. The stable stuff. Who you are, how you want to be spoken to, the rules of the house. Set it once and it applies to every chat.
Layer 2 is memory. The notebook your tool keeps on you while you walk around the building. Some of it gets auto-written, some of it you write on purpose, all of it is editable if you go look.
Layer 3 is projects. The conference rooms. Each room has its own rules posted on the wall (its own custom instructions), its own reference material on the table (its own files), and its own context that doesn't leak from room to room.
Layer 4 is files. The source material on the table inside each room. Brand guides, SOPs, past case studies, the authoritative stuff you don't want the tool to make up.
Most humans only ever use the prompt. The real unlock is context, then conversation.
Get this model in your head and the rest of the training starts to click. We're going to walk each layer.
Layer 1: Custom Instructions (The Lobby)
In plain English, custom instructions are the global facts about you and the rules you want your AI tool to follow in every single conversation. Your role. Your audience. How you want it to talk to you. Words it should never use. The format you prefer when it answers.
Custom instructions are available on every plan. Free, Plus, Teams, all of it. Settings, then Personalization, then Custom Instructions.
Before you even touch the text fields, check the base style and tone. You can dial it default, professional, friendly, candid, quirky, efficient, cynical. George rocks with candid because he wants direct and encouraging. Most humans never realize that lever is there. Same goes for the characteristics underneath: how warm, how enthusiastic, how heavy on headers and lists and emojis. If you don't want emojis, turn them down. You can add them in yourself if you want them.
Then there's the actual custom-instructions text. Two fields matter most: the "about you" field and the "anything else" field.
In the "about you" field, give the tool your name and your occupation. Don't just put "business owner." Stack the actual roles. For George, it's business owner, senior copywriter, HubSpot strategist, speaker, trainer. That stack tells the tool what hat you're wearing when you ask a question.
In the "anything else" field, here's the rule George uses (and the one that saves most humans from their biggest mistake):
Roughly 60% of your characters go to working context. Roughly 40% go to the rules.
Working context is the stuff that actually shapes the output: your audience, your voice, your constraints, who you serve. The rules are the guardrails: banned words, format defaults, length preferences (default to 150 words, expand only when I ask), things like no em dashes, use the word "flourish" instead of "thrive."
What burns most humans is the opposite split. They spend 80% of their characters on biography (where they went to college, every company they ever worked at, their kids' names). All of that can live in memory if you want the tool to remember it. It doesn't need to be in the lobby every single chat. The custom instructions are what the tool walks through every time. Make every character count.
Three Mistakes To Dodge In Custom Instructions
Three flavors of trouble George sees over and over:
Stale instructions. You're three jobs and 18 months removed from the version of you who first wrote these. The custom instructions still say you work at the old place. Update them.
Contradictions. Lines like "be concise" and "always include examples" fight each other. Pick a hierarchy. Tell the tool what to do, not just what to avoid.
No hierarchy. Default to 150 words. Expand only when I ask. Without that, the tool gives you the longest answer it can. You're not getting "deep," you're getting "wordy."
Here's the mini-homework you can do today. Open your custom instructions. Read what's in there. Delete what's stale. Run the SWOT-analysis prompt George shared on screen (he drops it in the chat, optimize-my-custom-instructions style) and rewrite from the audit.
Layer 2: Memory (The Notebook)
Memory is the notebook your AI tool keeps on you. It's a little bit of a database but not really a database. Some of it gets written automatically as you chat. Some of it you can ask the tool to save on purpose. All of it is editable if you go look.
Two pieces inside memory matter. Saved memories (the entries you can read, edit, and delete one by one) and reference-chat-history (the toggle that lets the tool reach back across past conversations). If you've ever typed "based on everything you know about me, create this thing," you were invoking the second one whether you knew it or not.
Here's the move most humans never make. You can explicitly tell the tool to save something. "Hey, save this SWOT analysis to my memory." Done. Live entry. Or you can flip it the other way: "Forget that, trash the entry I just saved." Both moves are one sentence away.
Why does this matter? Because memory is where the long-running facts about your work belong. The phrases you always use. The cliches you always replace. The format you prefer for YouTube packages. The structure you want every transcript turned into. George's memory has notes from almost two years ago that still shape every reply he gets.
Memory is the place the tool remembers you. Go in there. Read it. Clean it.
If you've never opened memory and looked at what's in there, today is the day. Settings, Personalization, Memory, Manage. Scroll. Delete the stale stuff. Notice what's there that surprises you. That notebook is shaping every answer you've gotten.
Layer 3: Projects (The Conference Rooms)
Projects used to be called custom GPTs. Custom GPTs still exist (more on when to use them in a second), but projects are the layer to lean on when the work is internal or just for you.
Two things about projects matter most.
Project instructions override your global custom instructions inside that project. That's huge. You can have a baseline lobby for every chat, then walk into a specific conference room with different rules posted on the wall.
Each project gets its own files. Last time George checked, you get around 5,000 characters for project-specific instructions plus a dedicated file area. Brand guides, SOPs, client-specific reference material, whatever the room needs.
Now here's the move most humans haven't found yet: Branch in New Chat. Once you have projects set up, you can thread them together.
Picture three projects. A research project (custom instructions and files all about how you want to research, what sources to pull from, what format to capture findings in). A writing project (custom instructions and files about how you want long-form content written, your voice rules, your structure). A social project (custom instructions and files about how each social platform wants its content shaped).
Start a chat in the research project. Do the research. Hit Branch in New Chat. All the context comes with it. Land in the writing project, turn the research into a blog article. Branch again. All the context still comes. Land in the social project, turn the article into platform-native posts.
Projects are the foundational base for the system you need to build. Stop thinking in tasks. Start thinking in systems.
That's the whole game. You're not asking the tool to do one thing one time. You're moving structured context through a series of rooms, each one shaping the work a little more, all without restarting the conversation.
Projects vs Custom GPTs (The Quick Call)
If the thing you're building needs to be shared with humans outside your org, still use custom GPTs. If it's internal (just you, just your team on a Teams plan), use projects. Sidekick Strategies runs internal work on projects and only ships custom GPTs when humans outside the company need to use them.
Layer 4: Files (The Source Material On The Table)
Files are the authoritative reference material you don't want the tool to make up. Brand guides, SOPs, past case studies, the documents that hold the truth about your business so the tool doesn't have to guess.
Files live at two levels. At the project level (every chat in that project has access to those files) and inside a chat (a one-time upload for one specific task). Most of the leverage is at the project level, because that's where the same reference material applies over and over.
The thing to watch for with files is hygiene. Three habits matter:
Don't go stale. If your brand guide changed six months ago and your project files still hold the old version, the tool is producing content from the old version. Swap it.
Cut the bloat. Three deep-research dumps from three different tools is a lot of context. Some of it is gold and some of it is filler. Cut what doesn't need to be in there before you upload.
Add yourself. AI research and your judgment are not the same thing. Before you hand a file to the tool, add the part only you know: your context, your stakes, your "here's how I actually use this."
We are the humans in the loop. The tool reads what we feed it.
The Build, Stacked
Four layers. One stack. One system underneath every chat.
- Layer 1, Custom Instructions: the lobby. Stable role, audience, voice, rules. 60% working context, 40% rules.
- Layer 2, Memory: the notebook. Auto-written and human-edited. Audit it, clean it, save on purpose.
- Layer 3, Projects: the conference rooms. Override global instructions, hold their own files, thread together with Branch in New Chat.
- Layer 4, Files: the source material on the table. Authoritative reference, kept fresh, trimmed of bloat, layered with your own context.
That's the 101. When you take these four and push them deeper, you start building what George (and most of the internet) calls a second brain. We even got a peek at his.
A Quick Look At Echo (What "Deeper" Looks Like)
Echo is the second brain Sidekick Strategies built for itself. Same four layers, just stretched further.
Custom instructions become identities. George B. Thomas has one, Sidekick Strategies has one, HubHeroes has one, every client has one. Each identity carries its own voice, audience, vocabulary, and rules.
Projects become folders. AI content by week, content by client, every workstream wrapped in its own context.
Files become an entire knowledge base of meeting transcripts, training material, frameworks, and source documents (we open it in Obsidian for the visual, and the graph view shows every file cross-linking to every other file).
Memory becomes a brain that knows what was decided last week and uses it to shape what you ship this week. When George asks Echo "answer real short, tell me how smart you are on HubSpot," Echo says "I'm an eight, I have Casey the HubSpot expert agent plus skills for CRM architecture, RevOps, workflow design, content ops, social, email nurture." That answer happens because the four layers got loaded the right way at scale.
When you use custom instructions, memory, projects, and files the right way, you never start from zero. You never start from garbage in, garbage out.
You're not expected to build an Echo. We're just showing you what's possible when the basics get taken seriously.
Your Homework Before Session 4
Two weeks. Three moves. Big payoff.
- Open your custom instructions. Read every word. Delete what's stale. Fix any contradictions. Establish a hierarchy. Run the SWOT-analysis prompt George shared on screen (and ask in the form below if you want it sent to you), then rewrite from the audit using the 60/40 split (60% working context, 40% rules).
- Audit your memory. Settings, Personalization, Memory, Manage. Scroll the entire list. Delete anything stale, off-voice, or just plain wrong. Save one new memory on purpose (any long-running fact about how you work) so you've used the lever once on purpose.
- Build one project end-to-end. Pick a workflow you run often (content for one client, your weekly newsletter, your sales-call prep, anything repeatable). Spin up a project. Write the project-level custom instructions. Drop in two or three reference files. Run one full task inside it. Notice how different the output feels with the context already loaded.
Bring it to Session 4 on May 29. We'll layer the prompt framework on top of the setup so every conversation lands clean.
Context Before Complexity
Stop thinking in tasks. Start thinking in systems. That's the shift Session 3 is built around.
Custom instructions are the lobby. Memory is the notebook. Projects are the conference rooms. Files are the source material on the table. Four layers, one stack, one system that lets your AI partner show up already knowing who you are, who you serve, and how you sound.
Once that stack is loaded, every prompt you've been typing for months starts producing dramatically different work. Not because the prompt got smarter. Because the context underneath it finally exists.
Whatever you build, ask yourself how it leaves them better than you found them.
That's the orientation. Build for the teammate who's drowning. Build for the client who's overwhelmed. Build for the audience you serve. And build for yourself, so you can augment the human at the center of all of it.
See you May 29 for Session 4, where we wire a prompt framework on top of the setup you've just built so the conversations you have with your AI partner start landing clean every time.
Your Next Move
Register for Session 4 (Free). On May 29 we layer a prompt framework on top of the four-layer setup so every conversation lands clean. Save your seat for Session 4.
See the Full 8-Session Series. Want the whole roadmap before you commit to the next live session? Every topic, every date, every recap. Explore the full series.
Ready to go deeper? The paid 4-week AI Content System training builds the real thing with you, with Claude Code and a complete identity-to-publish pipeline. See the paid training.
