In the age of AI agents and context windows a new valuable metric has been pushed to the forefront of my mind, this metric is time for a new developer (or AI agent) to onboard and contribute. Let me explain.

How quickly can you get going?

If I was to tear down your entire development setup on a project you work on, how long would it take you to get setup again? This means different things for different projects, for one person it might mean how quickly can I get my UI visible on localhost:5173 for others it might mean how quickly can I start draining messages from a Kafka topic.

As you do this, take note of any potential difficulties you have. Any potential gotchas! For example some of my Rust projects require the sqlx-cli installed to run DB migrations. Note that down. Codify it. Make it clear. Make it discoverable.

Now let's say this person is a new developer, let's assume mid-level experience profile in this tech-stack. How quickly do you think they could get going? Try it with your friends. Take note of any time they spend bashing their head against a wall. Add it to your todo list to resolve, or codify.

So, why is this important? AI agents are pretty good at writing code nowadays... I feel like that's no longer a hot take. The "problem" we have is context windows. Imagine if you and your co-workers had a similar setup to the show Severance but it was everytime you picked up a new ticket, issue, bug. That's exactly what Claude and Codex have to live through.

To get the most out of them, we have to keep this context as clear as possible. Long gone are the days of giving an agent your entire codebase compressed. Garbage in, garbage out. We gave them tools, like us as humans have. If I'm looking for my AuthService I don't read file by file till I find it, I grep, ctrl-shift-f, find and then work my way from there.

Steps to take

Firstly, definitely get setup with an AGENTS.md file, give it pointers on the structure. You can even AI generate an initial one and make edits. If you use pnpm tell it to use it, if you don't you're wasting context on telling it! I try to keep my AGENTS.md file as minimal as possible, giving it pointers on project setup, how to build, test and run any code or helper scripts I have defined.

Secondly, ensure your project has a clear structure. In my opinion in the AI age monorepos are a must, but if done badly they can really be a pain. If your project is large and made up of many subprojects, ensure there's a clear division between the services. Add a description of the structure to the AGENTS.md.

Remove any gotchas or tribal knowledge... this ones difficult and may require a bit of engineering effort. But trust me it's worth it in gained productivity and therefore saved cost down the line. I used to have a Python project where I had to depend on another repo via setting PYTHONPATH, I just know AI would struggle with the missing import unless I either... tell it. Or restructure my projects to consume the library as an actual dependency.

Some gotchas... are a pain to remove. But with all these tools at our disposal the cost of doing nothing has never been higher.

Finally, use the threads feature on these AI agents! When a task is complete, start a new thread clear the context. You don't have to ingest every ticket you've actioned today when picking up a new one. So please... stop forcing Claude to do it. This also has the added benefit of forcing you to break up larger tasks into subtasks. Which in turn makes code review easier! Win win!

Closing Notes

Short post of my thoughts, nothing overly formal I just like putting this methodology into writing. It's really improved my workflow with AI. If you haven't been able to tell I've recently done another deep-dive into vibe-coding and with Codex 5.3 and Opus 4.6... I'm really enjoying it.

That's all from me, any edits or notes I'll leave below!