Every productive AI interaction follows the same five phases.
This Is What My First Day Looked Like
Four steps. Thirty minutes. After this, my AI never started from scratch again. Yours won't either.
Create CLAUDE.md
2 minutes# My Project ## Tech Stack - React 18, TypeScript, Tailwind CSS - Supabase for auth and database ## Architecture - Feature-based folders - API routes in /api, components in /features ## Current State - Auth working, dashboard empty - Next: build the task list feature
You just gave your AI a memory. It'll read this every session.
Create TODO.md
3 minutes## Active - [ ] Build task list feature - AC: Users can create, edit, delete tasks - AC: Tasks persist in Supabase - AC: Drag-and-drop reordering ## Done - [x] Set up auth with Supabase
Your AI now knows what you're working on today.
Run your first session
5 minutesAI Response
I've read your project context. You're building a task management app with React 18 and Supabase. Auth is done. The next priority is the task list feature with create, edit, delete, and drag-and-drop reordering. I notice you're using feature-based folders. I'll create the task list under /features/tasks/ with a clean separation between the UI components and the Supabase queries. Ready to start with the data model, or would you prefer to begin with the UI?
That's the moment. Your AI just read your project without you explaining a thing.
Session 2, the compound moment
next dayAI Response
I see from PROGRESS.md that you shipped the task list CRUD yesterday. The drag-and-drop reordering from TODO.md is still open. Based on DECISIONS.md (DEC-002: "Keep UI state local, persist on blur"), I'll use React state for drag positions and only write to Supabase when the user drops. This avoids the flickering issue you noted in PROGRESS.md. Ready to tackle the reordering?
You didn't tell it any of that. It read it. That's the compound effect.
That's it. Four files. Thirty minutes. I did this over a year ago and haven't re-explained my project since.
Four Habits I Use Every Day
These aren't abstract ideas. They're four things I changed about how I work with AI, and they reinforce each other.
Plan in Markdown, Not Jira
How it was
You keep tickets in Jira and context in Slack. Your AI can't read either, unless you've given it MCP access.
How it is now
Move the plan into the repo as TODO.md and your AI reads it every session. No board, no grooming, no re-explaining what you're working on.
Your AI reads the plan because it lives where the code lives.
Protect Decisions, Not Code
How it was
You guard your code because rebuilding feels expensive.
How it is now
Here's what's funny. With AI, building is the cheap part. The wrong architecture is what kills you. DECISIONS.md and CONSTRAINTS.md protect the choices that are actually expensive to get wrong.
I've rewritten a component in an afternoon. A wrong data model followed me for sprints.
Align with Files, Not Meetings
How it was
Your AI drifts from your goals. You catch the violation three sessions later.
How it is now
Point it at your PRFAQ, principles, and constraints before it touches a line of code, and every session starts aligned. The drift stops.
It can't drift if it reads the destination every time it starts.
Measure Compound, Not Output
How it was
Every session starts with re-explaining the project. Same effort, same output. The flat line.
How it is now
Now today's session builds on yesterday's. PROGRESS.md carries what happened, DECISIONS.md carries why, and the flat line bends into a curve.
If your sessions compound, you're moving. If they re-explain, you're burning tokens.
These reinforce each other. Markdown planning feeds alignment. Protected decisions enable compound sessions. It works because it's one loop, not four separate habits.
Marek is sharing these while they're working, so you can skip the months of figuring them out.
Three Protocols I Use
I use different protocols depending on the task. Here's how they break down.
uno
Operate
AI works FOR you. Clear task, trust output.
- → I use it for trip planning
- → Documentation projects
- → Research tasks
Habits 1 & 4: plan in markdown, measure compound
Get Template →duo
Construct
AI works WITH you. Iterative, role-switching.
- → Code projects
- → Product builds
- → Team handoffs
All 4 habits: planning, decisions, alignment, compound
Learn duotre
Automate
AI works AMONG systems. Quality gates, approvals.
- → Pipelines
- → Compliance checks
- → Automated workflows
Habits 2 & 3: protect decisions, enforce alignment
Coming soonNot sure? Here's the simple version: If you write code, start with duo. If you don't, start with uno. You can always switch later. The files are the same, the protocol is different.
One Cycle, Three Protocols
The atomic cycle is the same everywhere. What changes is scope. How much context you load and how AI collaborates with you.
AI Basics 101
Start here
The memory myth, files as ground truth, hallucination prevention, and token economics.
Read the basics →uno / Operate
AI works FOR you
Trip planning, documentation, research. Habits 1 & 4: plan in markdown, measure compound.
Learn uno →duo / Construct
AI works WITH you
Software projects, product builds. All 4 habits: planning, decisions, alignment, compound.
Learn duo →tre / Automate
AI works AMONG systems
Pipelines, quality gates, approval workflows. Habits 2 & 3: protect decisions, enforce alignment.
Coming soonThe 4 Levels
From Jet Ski to Direct Intent
Most people explore with AI before directing it. Learn the 4-level model: when to build jet skis, when to add the strategic roof, and how to graduate from exploring to executing.
Read the guide →AI Readiness Check
DIY maturity assessment
12 production-ready AI capabilities across 4 categories. Check off what your team has solved. The result tells you whether uno, duo, or tre fits your organisation.
Take the assessment →Notes from running the protocol.
Working notes on what the protocol catches as the platform evolves. Each article continues the thread.
Insight · Constraints · Latest
The word list was never the product →
We thought we were building a spelling app. We were writing down a theory of how a child learns, and the constraint that froze everyone onto one list just lifted. Building the fence, not the net.
Insight · Origin
Extracted, not designed →
The duo protocol was never designed up front. I built a gym app for 75 commits, then read the patterns back out of the git history.
Insight · /goal
How /goal slots into the loop →
Claude Code's /goal swaps a different model in as the grader. Why that maps to the protocol's RECORD step, and where Codex's budget_limit fits in.
Position · Claude Code
Why we double down on Claude Code →
Five primitives the protocol has always argued for. Five primitives Claude Code now ships natively. They line up because they're modelling the same thing.
Insight · v1.3
Why v1.3 is a minor bump →
Five layers of the protocol moved in the same release. None of them was a bug fix. The brewery-and-recipe model (protocol-level rules vs flavour-level rules), and why we called it minor.
Not Sure Which Protocol?
Do you need AI to write code?
- → Yes → duo (Construct)
- → No → What kind of work?
New to all this? Start with AI Basics 101. It takes 10 minutes and covers the fundamentals.
Ready to Go Deeper?
The duo/ab protocol is the Construct-mode protocol for software development. Clone the template, follow the guide, and start compounding.