KAYGEE

Two Monday Mornings

~/my-project

Claude Code v2.1.101

Opus 4.6 (1M context) with max effort

> Let's get started

Is this you?

> I'm working on a fitness app with offline sync and role-based auth. Here's what I need you to know...

> No, we decided against Next.js three weeks ago. I told you this yesterday.

> Forget it. I'll write it myself.

Same briefing. Same corrections. Every morning.

This could be you

> Loading playbook... CLAUDE.md, TODO.md, DECISIONS.md

Morning. You shipped the auth refactor Friday. TODO.md has the dashboard migration next. Want to start there, or did the weekend change your thinking?

It remembered. You didn't have to.

The difference is a few markdown files your AI reads before every session.

Learn the Protocol
Read the story →
What Surprised Me

Why Files Work

AI forgets everything between conversations. Start a new chat, and it's meeting you for the first time.

But files don't forget.

When you write what you learned in a file, AI reads it next session. When you track what's done, AI knows where you left off. When you log what didn't work, AI doesn't suggest it again.

Here's what that feels like. You open a new session on Monday morning. Instead of "What are we building?", your AI says "I see you shipped the auth refactor on Friday. Ready to tackle the dashboard migration from TODO.md?" That's not magic. That's a file it read 200 milliseconds ago.

Your markdown files ARE your playbook.

→CLAUDE.md What this project is and how it works
→TODO.md Prioritised task list, so nothing gets lost
→DECISIONS.md Why we chose this approach, so we don't re-debate
→PROGRESS.md What happened, so the next session knows
→PRINCIPLES.md Distilled wisdom that compounds across sessions
→CONSTRAINTS.md What we rejected and why

Each file solves a specific forgetting problem. Together, they're your playbook.

I set these up once per project. About 30 minutes. It's been running for over a year and every session is still building on the one before. You're not writing documentation. You're teaching your AI how to think alongside you.

The Compound Effect

Without the protocol, every session resets. With it, every session builds on the last.

Without Protocol

Session 1
Same
Session 2
Same
Session 3
Same
Session 4
Same
Session 5
Same

Equal effort, equal output. Every session starts from scratch. AI forgets your decisions, patterns, and context.

With Protocol

Session 1
Setup
Session 2
Growing
Session 3
Building
Session 4
Compounding
Session 5
Flying

More effort initially, then it compounds. AI remembers your decisions, checks your patterns, and builds on what came before.

1.

Write decisions down

Markdown files in your Git repo capture decisions, constraints, and patterns.

2.

AI reads your playbook every session

Context loads instantly. No re-explaining, no drift.

3.

Each session builds on the last

New decisions get captured. Context grows. The protocol compounds.

Where the protocol came from

Extracted From 400+ Commits.

The Gym Workout Tracker is the full-stack app the duo/ab protocol was extracted from. There was no protocol when it started. Each session read what came before, and wrote for what came next.

400+

Git commits

20

Architecture decisions

1

Extracted package

Live

Deployed & used daily

Moments that changed how I build

→

The Fix Chain

Protect decisions

That feeling when your AI says "this fix will break 5 other things" and it's right. Not because it's smart. Because DECISIONS.md told it how the architecture connects. It caught 6 cascading issues from one routine bug fix. I would have found them over weeks. It found them in seconds.

→

20 Decisions Into 5 Principles

Compound sessions

Sixty days in, I asked the AI to review my architecture. It synthesised 20 separate decisions into 5 governing principles in one sitting. It remembered conversations I'd forgotten. It connected patterns across sessions I'd never linked myself. That's 60 days of context, working for me.

→

Auth Became Its Own Package

Alignment

Twenty-two commits, cleanly separated. No errors, no missed dependencies. The AI knew exactly where the boundary was because I'd written it in CONSTRAINTS.md three months earlier and never thought about it again. Once you have the compass set, it follows it without more prompting.

What went wrong

The protocol doesn't prevent mistakes. It prevents the same mistake twice.

×Weight tracking UX failed. First implementation was confusing. Had to rebuild. But PROGRESS.md logged what didn't work, so the second attempt avoided the same patterns.
×Protocol files grew too large. After two months of context, AI started hitting token limits. Led to a hygiene system, archiving old progress, consolidating decisions. Now a core part of how the protocol works.
Try the Gym App

Built by one person and one AI who never forgot a conversation.

Watch It Work

One person and their AI building something real. Three minutes, no slides, no theory.

Want the full walkthrough with edge cases?

Watch the 10-minute real-time demo
Insights

Notes from running the protocol.

Working notes on what the protocol catches as the platform evolves. The newest essay, then the full series.

Insight · Origin · Latest

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.

Read the full series →
🤔

Not sure which protocol fits?

Take our 1-minute quiz to discover whether uno, duo, or tre matches your workflow.

Find Your Protocol
Common Questions

Everything You're Wondering

Questions from people building with AI

Still have questions? get in touch.