Michael Gardon
CHIEF AI Accelerator
Phase 3 — Set Up To Work · Pre-work for Module 3

Working with Projects

How to give AI persistent context so it stops starting from zero every conversation — and why the right project setup determines whether Ring 2 actually pays off.

15 min read · Bring 1 real project to the live session
The Problem

Auto-memory remembered your name. It didn't remember the memo.

Every tool handles memory differently. ChatGPT has Memory that auto-summarizes facts about you across chats. Gemini saves personalization info. CoPilot grounds in your M365 graph activity. Claude, by default, has no user-level memory at all.

So why doesn't this solve the context problem? Because the memory each tool gives you for free is shallow, automatic, and not fully under your control:

The result: you still spend the first ten minutes of every important conversation explaining your organization, your audience, your constraints, your tone. Next week, same problem, different angle — and you mostly start over.

Relying on auto-memory for serious work is like hiring a skilled consultant who remembers your name and rough industry but nothing else — every morning.

Projects give you the deep, explicit, controllable context layer that auto-memory cannot.

Why This Matters

Every ring expands the context AI has access to.

Ring 1 lived entirely in the Conversation layer — every prompt was self-contained. Ring 2 begins when you expand into Project and Canonical layers. A project (whatever your tool calls it) is the container for both. Without it, Ring 2 has no home.

Layer 01
Conversation
What you paste or type into this chat. Ephemeral. Dies when the chat ends.
Ring 1 — The C in 4 C's
Layer 02
Project
Persistent context for this work area. The current initiative, the relevant files, who's involved.
Ring 2
Layer 03
Canonical
About you. Your voice, taste, experience, operating principles. Travels across all your work.
Ring 2
Layer 04
Integrations
What AI can read, do, and where outputs go. Connections, tools, scheduled triggers.
Ring 3 — Future
Today: install layers 02 and 03 Eventually: layer 04 →

This is why the next move is project setup. Not because projects are cool. Because encoded judgment has to live somewhere AI reads automatically — otherwise you'll keep typing the same context every chat, and nothing accumulates.

The Anatomy

Four components, regardless of tool.

Different tools call these different things, but the underlying structure is the same. Miss any one and the project becomes a folder of chats, not a working system.

01

Instructions

Your standing brief

Written once, active in every conversation inside the project. Who you are, what this project is for, how you want AI to think with you. The difference between a generic assistant and one that already knows your context.

02

Knowledge Files

Your reference library

Documents AI reads before responding. Strategy memos, brand guidelines, previous analyses, your voice and taste docs. AI references these directly — you don't paste them into every conversation.

03

Persistent Context

What carries between chats

Decisions, directions, conclusions across time. Some tools handle this automatically; others require manual updates to knowledge files. Either way, the project is where work continuity lives.

04

Conversation History

Your work archive

All conversations within the project in one place, searchable. Return to any thread, pick up where you left off, or trace how a decision evolved over multiple sessions.

The Vocab Map

Same concept. Different buttons.

Each major tool has its own name for projects. Knowing the map lets you apply the same discipline regardless of which tool you use.

Tool What it's called Default behavior
Claude Projects Explicit. You load files, write instructions, manage what AI sees.
ChatGPT Projects (newer) or Custom GPTs (older) Mixed. Memory is auto across chats; project knowledge is manual.
M365 CoPilot Copilot Studio Agents, Notebooks, or saved prompts Implicit. M365 graph grounds across your work email/files automatically; explicit scoping is optional but powerful.
Gemini Gems + Drive grounding Mixed. Workspace data is grounded; Gems add explicit scope.
The Philosophical Split

Explicit vs. implicit context.

Each tool makes a different bet about how much you should do yourself. Knowing which bet your tool made tells you exactly what work is still on your plate.

Claude

Explicit by design

Assumes you'll load context yourself. Nothing carries between projects unless you put it there. No surprise memory, no ambient grounding.

Cost
You actively maintain the project.
Benefit
No drift. You control the cabinet completely.
M365 CoPilot

Implicit by default

Grounds in your M365 graph — email, OneDrive, SharePoint, Teams, Calendar. You don't create a project layer; your work environment is the project layer.

Cost
Less control. Broad, sometimes noisy grounding.
Benefit
Massive head start on context.
ChatGPT

Mixed model

User-level Memory (auto), Projects (manual), Custom GPTs (shared scope). The most layered model — most flexibility, most decisions to make.

Cost
Easy to lose track of what AI knows and how.
Benefit
Choose how much auto vs. manual.
No tool gives you Ring 2 for free. The explicit-context discipline matters more than the tool choice — your AI will not sound like you until you've encoded what you sound like.
For CoPilot Users Especially

Even with M365 grounding, you still need projects.

The "I have CoPilot grounding, I don't need projects" instinct is half-right and creates four real problems.

  1. Discoverability noise
    Grounding pulls from ten years of files. AI might reference an outdated 2022 strategy instead of the current one. No authority signal tells it which to trust.
  2. Staleness
    Old drafts, abandoned plans, superseded versions all live in the same graph. AI doesn't know which is current.
  3. Authority conflicts
    Multiple docs say slightly different things. AI guesses which one wins.
  4. Personal vs. organizational confusion
    Your personal taste and voice don't belong in the company SharePoint. But CoPilot can only read what it has access to. Personal canonical content ends up in awkward limbo.
M365 grounding is your floor, not your ceiling. The work of Ring 2 is building a narrow, curated lens on top of broad grounding — because broad grounding is noisy when judgment matters most.

A scoped container — a Copilot Studio Agent, a Notebook, or a designated OneDrive folder you point AI at — solves all four problems. You still need it.

The Decision

When to build a project — and when to skip.

Not every task needs a project. The question: does AI need to know something about you, your work, or your context to give you something genuinely useful — and will you need that same context more than once?

Build a project

Ongoing, context-dependent work

  • You return to this problem over days or weeks
  • AI needs to know who you are to give useful answers
  • You have documents or constraints AI should always know
  • Each conversation should build on what came before
  • The work produces artifacts worth keeping in one place
The test: if AI already knew everything about this problem, how much better would the answer be? If significantly — a project is worth the fifteen minutes to set up.
Beyond Scoping Work

What projects actually become as you mature.

Projects start as folders for related chats. They become the container for everything Ring 2 and Ring 3 you do.

One mental model for agents: an agent is a project with action permissions. The project is the trust boundary.
Before the Live Session

Pick your project.

Identify one real project you'll set up live with us. It does not need to be your most important work. It needs to be something you return to weekly, where AI knowing your context would change the quality of the output.

Examples that work well

Bring three things to the live session

01

The project area

One sentence: what is it? "I advise mid-market PE portfolio companies on operations." or "I write a weekly insurance industry newsletter."

02

One document

The most context-loaded artifact you have for this work area. The first thing AI should always know.

03

Three repetitions

Examples of context you typed more than once in the last week. The most important item — tells us exactly what should live in the project.

Why the repetition examples matter

The repetition examples tell you exactly what should live in the Project layer (context for this work area) versus what belongs in the Canonical layer (about you, across all your work). We use them as the seed for your install.

The Live Session

What we'll do together.

  1. Anchor — Why Ring 2
    10 min
    Surface the repetition examples from the room. Make the pain visible before showing the cure.
  2. The Four Context Layers
    10 min
    Where context lives. The Ring 2 skill is knowing what goes where.
  3. Vocab Tour
    10 min
    What "project" means in your specific tool. The map between concept and button.
  4. Live Install Demo
    15 min
    Watch an end-to-end setup. Project created, files loaded, capture skill installed.
  5. Hands-on Install
    15 min
    You set up your own project, in your own tool, on your own work.
  6. First Capture
    15 min
    Run a real prompt in your new project. Type wrap. See your first FOR THE FILE output. Paste into signals.md.

By the end, you walk out with a working project containing your first knowledge files, the Capture Skill installed, and one captured signal. That's the install of Ring 2. The real work begins the next morning — using it every week, watching it accumulate, watching your AI get smarter about you on its own.

A Note on Tool Specifics

Snapshot, not permanent map.

Specific interface details and feature names change. The underlying concept — giving AI persistent context to eliminate the re-explanation tax — does not. When something described here doesn't match what you see in your tool, check the support documentation for that tool. The principle is durable; the buttons evolve.

On memory features specifically

AI tools are rapidly evolving how they handle memory. Claude has been adding memory features in some surfaces. ChatGPT's Memory expands periodically. CoPilot's grounding behaviors continue to develop. Gemini's Saved Info works differently on consumer vs. Workspace accounts. Treat the specifics in this document as a snapshot, not a permanent map. The underlying principle — that projects give you deep, explicit, controllable context that auto-memory cannot replicate — holds across the changes.