I run a retro skill at the end of my sessions with Claude. Not once in awhile, every single time.
If that sounds like it's a bit much, I understand. But I really feel like it's the biggest thing I can think of that has increased my productivity with Claude. More than Sonnet versus Opus or any MCP server yet.
How it works is that at the end of every session regardless of activities I have Claude write a short retro. It lists out what was shipped, what decisions were made, what issues cropped up and what is next. The next session will read the top of the notes first as part of session start protocol.
I try to keep it that simple, keep it short and sweet as my Agile friends would advise.
But here is what happens over time. The retro isn't just a log, it feeds a system of active actions and modifies the rest of the memory system — and this affects every session going forwards. Mistakes that are caught once are prevented in the future, similar to root cause analysis and action items when a system has an issue at work. Patterns I like or that make me more productive are reinforced. The agentic loop tightens.
I'll give you one of my examples. Early on, Claude was making edits to my files without showing me what it was about to change. It would just write it and commit. I caught it and said hey, stop and propose what you are about to write. I need to confirm then. That went into the active actions list. Now every session it shows things for approval allowing me a second to think about it. One correction to fix an annoying problem.
This is what compounding context looks like. It's not the AI model changing, it's just the system around it getting better. A lot of these decisions from the retro end up becoming memory files on their own and patterns that make me more efficient. Each session becomes a small layer of improvement. Over time, they stack and watching things improve is exciting!
I see a lot of people outside IT treat AI like a stateless tool similar to how they use Google. Every session is a fresh start. But when you're building anything at all that leaves the best parts out. The system can remember so much.
One of the more senior developers on the team that I lead, Jon Kloss, took the concept of a sprint retro and turned it into a full enforced developer workflow. It logs incidents when Claude makes mistakes, structured entities with categories. It has a dedicated retrospective skill that analyzes patterns and drafts edits.
It's a whole pipeline: design through Socratic questions, generate specs, verify, retro it all. The retro is a first class citizen in the system and it gets better over time. I'm so proud of how creative and smart he is, and he built this while working on a massive project late each night. He's that dedicated.
His repo is public if you want to see a fully engineered version of retros in development workflow and throw him some stars!
My own private patterns are simpler but the concept is the same. The session doesn't end when we ship. It ends when we've captured what happened so the next session is better.
Out of the box AI tools are great and I'm having a blast. But the compounding concept which makes session fifty so dramatically better than session one, that comes from tightening the loop. Feeding it back in.
Discipline is king, and reflection keeps me getting more productive.