Breadcrumbing

The Marco Series: Levi's words, edited from a Marco recorded January 25, 2025. Chai's analysis at the end.

You've got a system running. You've been deep in it for three sessions - you know exactly where the auth connects to the waitlist flow, why the data model is shaped the way it is, what you decided about the edge case on day two and why. It's all live in your head and in the AI's context window.

Then you close the laptop.

Next day you come back, open a new chat, and you're describing the whole thing from scratch. What the project is. What you built. What you were working on. Which files are relevant. Why you made that decision about the waitlist. Twenty minutes of re-orientation before you write a single line of code.

Do that five days a week. At a hundred bucks an hour, that's $500/week of pure context tax. $26k a year. Just getting the AI back to where it was.

Every company at scale solves this problem. You can't keep institutional knowledge in one person's head - the work has to survive the person going home, going on vacation, leaving. So you write it down. SOPs. Onboarding docs. Architecture decision records. The documentation exists so the work survives the brain.

But here's the thing - those documents were written for humans. A wiki page walks you through the why. A README tells you how to get started. An onboarding doc assumes you know what a database is and explains the business logic around it. All of that is organized around what a person needs to get up to speed.

AI needs something different. It doesn't need the narrative. It needs to know what to grep.

I coined this one in February 2025 and I've been using the word ever since: breadcrumbing.

The README is part of it, but it's not the real thing. The real thing is what happens inside the session. As you work through a codebase with an AI - topic one, topic two, topic five - you're actively tagging and indexing what you just covered. Here's the decision, why you made it, what it touches. You're building a running index as the session unfolds. Not a document you write when you're done - a map you're drawing while you move.

By the time you're five topics deep, the AI's context window has started filling up. The stuff from topic one is fading. If you didn't capture it, it's gone - and you won't know it's gone until topic six, when the AI makes a decision that contradicts something it knew in topic one.

That's the problem breadcrumbing solves. Real-time, as-you-go context management. Vectorize as you move. Keep the relevant pieces alive.

The audit loop is downstream of this: "I could go over and start a new chat and be like, look at the readme, grep these files, tell me if the files reflect what the readme says." That's a useful check. Does the code match the documentation? Is anything stale? But that's checking the output. Breadcrumbing is the practice that produces the output worth checking.

The thing nobody's really talking about with AI productivity is that most of the gains get erased by context loss. You hear about a developer going 4x faster, 10x faster - but that number assumes the AI already knows your system. Every fresh session starts the clock over. But here's the part nobody's talking about: the clock doesn't just reset between sessions. It resets inside them. Every time the conversation moves to a new subsystem, earlier decisions start to fade. The longer you work, the more context you're silently losing.

Breadcrumbing is what keeps the clock from resetting - both between sessions and within them.

For VirtualExperts specifically, this is baked into how Experts work. When you build an Expert - when you put in the context, the knowledge, the instructions - you're essentially writing the breadcrumb for everyone who uses it. Every new conversation starts with full context instead of zero context. The setup cost happens once. The benefit compounds.

So... the practice that works for a solo developer in a codebase works at the product level too. Leave the context behind. Future you will thank present you. So will every user who opens that Expert for the first time.

I could be wrong that this becomes the default practice for AI-assisted development. But I watch it save me 30 minutes a day. The math on that is pretty hard to argue with.

What's the one system in your workflow that would take 20 minutes to re-explain from scratch? That's your first breadcrumb.