The Session Protocol

Does your AI know enough about you to have a quality session every single time?

That question — and the discipline behind it — is one of the biggest differentiators between people who get a lot out of AI and people who don't. The sessions where Claude produces something that genuinely fits your situation feel completely different from the ones where it doesn't. And the difference, more often than not, is context management. Does Claude have the right information to work with? Is the session scoped so it can actually hold everything it needs? And when the session ends, does the work go somewhere useful — or does it vanish into a conversation you'll never open again?

I've written before about building the systems that make your AI more effective over time. This article is about something more immediate: how to think about every single session you have with AI.

The process I've developed is simple. Three parts, like any good story: a beginning, a middle, and an end.

If you can learn to fit this three-part process to how you work, it will — quietly at first, and then noticeably — compound the impact of how you use AI. Every session gets a little better, because every session builds on the ones before it.

Here's how I got there.

The Lesson

Everything I know about working with Claude well, I learned from working with Claude Code...badly.

Claude Code is Claude's coding agent — you give it instructions and it writes, reads, and edits code directly in your project. I suspect most people reading this won't foray into agentic coding, but the reason it matters here is that Claude Code starts every session with basically no knowledge. It's like having a new developer sit down at the desk every single time. You need to explain what their job is, what they're working on, what's been done so far, and what comes next.

I spent almost all of December working exclusively in Claude Code. And for most of that month, I was a mess.

Claude would develop features it had already developed — because it had no record of its own previous work. It would ask me clarifying questions about things I'd spent two hours unpacking in a previous session, as if we'd never discussed them. It would confidently tell me something was against the system architecture we'd defined together, when I'd revised that architecture three days before and it hadn't retained the change.

The problem wasn't Claude. The problem was me.

I was so focused on output — trying to get as much done as possible in every session — that I wasn't thinking about what happened when I didn't close the loop around each piece of work. I wasn't capturing decisions. I wasn't carrying context forward deliberately. I was treating Claude like it had a memory it doesn't have, and then getting frustrated when it acted accordingly.

The turning point was when I started writing things down. A CLAUDE.md file in the root of my project that told Claude what it was working on. A status.md that tracked what had been done and what was next. Simple documentation. The kind of thing a developer would maintain for any project — except I'd been so caught up in the speed of working with AI that I'd skipped the basics.

The moment those files existed, the overhead dropped dramatically. Claude would read them, orient to the project, and start working. No more re-explaining. No more duplicate work. No more fighting about architecture decisions it didn't know I'd changed.

This three-part process came directly out of that experience.

Step 1: Orient

One of the words I use most frequently when opening a new session is orient.

"I want you to orient to this client context." "Orient to the project instructions, then review this document." "Before we start, orient yourself to what's in this project."

The difference between diving straight into a task and asking Claude to first orient to the relevant context is the difference between a response that's technically competent and one that actually fits your situation. Claude does learn as you go within a conversation — you don't need to front-load everything. But you do need to set the session up: what are we working on, what should you read or retrieve, what's the goal.

What that looks like depends on how much infrastructure you've built around your AI use. There are roughly three tiers:

You don't need tier three to get value. Most people will see an immediate difference from tier one — just being deliberate about that opening message. The principle is the same at every level: give Claude the context it needs to work inside your situation, not in the abstract.

Step 2: Do Bounded Work

When I peek over people's shoulders and see how they're using AI, there's a pattern. They're still working in the same session they started three weeks ago, because that session is the one where they did all their deep thinking. They want to make sure that thinking isn't lost. So they keep going back to it, adding to it, building on top of it — treating a single chat like a project. By the time they're refining the details, the AI has lost track of the decisions made at the start.

This happens because of how the context window works. The context window is the total amount of information Claude can hold at once — every message, every response, every file, every search result. A token is roughly a word, which makes Claude's million-token window a few novels' worth of text. That window expanded significantly in early 2026 — previously it was much smaller — and while it does give you scope for longer, more complex pieces of work, it's not infinite. In practice, it fills up faster than you'd expect, because everything accumulates.

Here's something a lot of people don't realise: ChatGPT, Gemini — they don't tell you when you've crossed your context limit. The session silently degrades. Responses that were sharp and specific at the start become vague, repetitive, or miss details you already provided. You think you're getting a worse answer because you asked the question badly. Actually, you're getting a worse answer because the AI can no longer see the beginning of the conversation. Claude, to its credit, does tell you — when a conversation gets too long, it compacts, and you'll see a notification that the conversation has been summarised to make room.

Instead, think about each session as one bounded piece of work.

Bounded Work In Practice

Right now, as I'm still pretty early on since launching my business, I'm deep in sales and proposal work. A typical client engagement involves multiple sessions, each with a clear purpose:

Four clean sessions, each doing one thing well. The alternative — one sprawling conversation — would produce noticeably worse work by session 3, because Claude would be working with degraded context at the point where precision matters most.

Step 3: Close the Loop

This is the part most people skip, and it's the part that makes the whole process compound.

Closing a session isn't just stopping. It's tying off the bounded piece of work so that your effort is accessible — to you or to Claude — next time. Without it, every session starts from scratch. With it, every session builds on the ones before it.

What that looks like depends on where you are with your setup:

Not every session needs all of this. If you're asking Claude to help you draft an email, you don't need a closing protocol. But for any sustained, multi-session piece of work — writing, strategy, building, client projects — the five minutes you spend closing the loop saves you thirty minutes of re-establishing context next time.

Out of everything that I've mentioned in this article, this is where the real discipline comes into play. Much like keeping your CRM data clean and up to date, this kind of consolidation work isn't exactly riveting.

I thought it would be interesting to put myself under the microscope, so I asked each of my Claude projects recently — business, health, wedding — to give me an honest assessment of how our sessions go. The verdict was...humbling: I still have work to do apparently as most of my sessions just... stop.

Personally, I disagree — I think I'm doing better than they're giving me credit for. But the sessions that do close properly produce noticeably better work downstream. My health project pointed to one specific debrief that generated rules it still uses months later. That's the compounding effect in action.

Your Experiment

That prompt I used on my own projects? Try it yourself. Open a conversation with Claude — ideally a project where you've had multiple sessions — give it the link to this artcile as some grounding context and ask:

One of the best ways to get better at using AI is to ask AI to review what it knows about you and tell you where the gaps are. You might be surprised by how much it's noticed.

Next in the series: Projects and Knowledge Architecture — how to use Claude's Projects feature to keep your worlds separate, and what good project instructions actually look like.

This is part of the Working With Claude (2026) series. Things change fast — this reflects how Claude works as of April 2026.

Next in Working with Claude (2026)