A Saturday morning with Claude Code
This Saturday morning I woke up and just wanted to code.
April 19, 2026
LLM coding has come a long way in a few years. Two years ago it was great at writing boiler plate and small one shot features. Having it contribute to a larger project was difficult and often perilous. Quality was very hit or miss, lots of "slop". Minor features would cause it to rewrite swaths of code, and you had to micromanage it aggressively. Last year brought really smart models like GPT4.5 which could produce good results but still required a lot of steerage if you were working on a large or complex project. This is now changing rapidly with Claude Code and the other latest frontier models and harnesses like Gemini/Antigravity and GPT-5.x/Codex.
For me Opus 4.x (4.7 as of time of writing) is genuinely game changing. It writes quality software and can contribute to large projects effectively. I now find myself thinking about writing software when I'm trying to fall asleep. I now have this super power letting me build things almost as fast as I can think of them. Quality things no less. I am the bottleneck as I context switch between several parallel agents to review their code and plans. What this means for the future of engineering and technology org design I have not quite figured out yet, mostly because capabilities grow month by month.
This Saturday morning I woke up and just wanted to code. That's been more and more common for me lately. I feel so inspired to just build and ship software. Even before a cup of coffee I was adding features to our internal clinical-design support tool Athena, performing GCP and AWS infrastructure audits, and building a new feature for a product we're launching in a couple weeks. All of this performed rapidly and effectively. I even got a hike in before lunch.
As I get better at my Claude Code work cycle, I thought I'd jot down a few of the things that have worked really well for me:
- Start with planning. Build an spec document together. Build extensive documentation on design choices. Have an idea how you want to host/run your app. Claude's intuition here is often poor, but it can implement basically every schema under the sun (want an ECS cluster in AWS? Pure severless GCP CloudRun? Complex Terraform? No problem!) Then have Claude actually start writing code.
- You need to understand your own security model. Claude will make suggestions but its up to you to really understand it and ensure Claude implements against it and implements tests to verify it before anything gets merged/released. Tools like Claude Code will get better at this with time, but for now you really need to be in the drivers seat here.
- Have Claude build tests as you go. Every endpoint, every feature. It's generally good at this but watch that it doesn't forget down the road. Make sure to remind it after each major feature to ensure test coverage is good.
- Everyone has their own prompting style, but I find verbosity wins for me. Give it as much context as possible. Speak to it like an engineering manager writing tickets for a junior engineer. (You can wire it into your ticketing system directly too) Human language, full sentences. LLMs work best when the input tokens better match what's in the training dataset.
- Build verbose logging into whatever you are building. Enough that you can verify that an app is working properly because...
- Tell Claude how to troubleshoot what it builds, give it read-only database and log access. Push it to add logging so that it can troubleshoot whatever its building. Then make it check the efficacy of new changes. Have it doing QA on live requests once a feature ships. This feedback loop is great because LLM's performing frequent log review is relatively cheap.
- Make Claude document the development and troubleshooting process as it does these log / database reviews. Ask it to document it for you and for it. The combination of writing instructions for human and LLM agents will get you better results when refreshing the context for future work, especially about how its discovering and squashing bugs, as well as your testing and security model.
Our team at Origin has embraced Claude in a big way. What tips and techniques have you found that works well? Shoot me a note!
Happy building!