Some notes and observations from using llms

I've been using some form of LLM code assist since GitHub Co-Pilot entered beta. Over the years that's included various GPT, Gemini and Claude versions along with trying different editor and CLI based assistants. How I integrate and use these tools is something that continues to change for personal and professional projects, but I thought it might be time to write down a few thoughts.

• This one should be obvious, but in case it's not you need to try different models with different projects and teams to find which is the most efficient in that context. Right now I've have found Claude Sonnet 4 to be the model that provides the code I prefer to maintain, but it helps to have evaluation criteria that you use to check the results of each new model release, and for testing models across different vendors (Anthropic, OpenAI, Google, etc).
• Fundamentally these tools can change how you work if you let them, but:
• It's particularly important when using these tools to consider how different "modes" (agent, agent + mcp, ask, etc) trade off agency and how that impacts yourself, and your team.
• For many of us code is only part of the project, and sometimes not even the hardest part, but consider that the LLM is more than happy to generate pages of code. That can be a burden to maintain, and if operating in agent mode I will wager when the agent runs off and implements a full feature that you don't know it the same way you would if you used the LLM to provide suggestions/ snippets and or you implemented it yourself. There is a cost there, short and long term.
• There is a larger team and organization conversation going on right now that you need to engage in. These tools can boost and harm moral, and strong teams are foundational to success short and long term.
• I go back and forth between how I use agents depending on the structure of my week. Some weeks I'm spending a lot of time pairing, meeting or doing task that don't provide hours at a time to be head down focused on the code in a flow state. When that happens I'm happy to have tools like LLM agents that I can prompt and check back on, but "with great power comes great responsibility". Don't check in that code without doing a thorough review on your end first (including a very detailed review of any generated tests). Your team will thank you for it, and you won't find people (like myself) leaving frustrated review comments about generated nonsensical code.
• LLMs are very verbose, tame this with your prompts if you expect to thoroughly review and integrate their output.
• Be careful when allowing LLMs to generate tests. I've heard so many people talk about how great it is they can use these tools to write test for code that doesn't have any, or to have it write test with the new feature they are implementing. That's not wrong, LLMs can generate test, but just like building maintainable systems is a skill writing good test is also a skill. I would wager that the code training datasets are not curated to projects such as AOSA. We would probably disappointed to see where most projects sit on a quality curve for training data sets. I can personally share that time and again LLMs will generate test that erroneously pass. They will remove assertions, write assertions that never fail, or write test that never call the function or DUT. Yes, you can have an LLM generate good test, but it takes work, and be cautious, our test are one of our strongest signal points in a project. Failing to keep their integrity strong will only lead to bugs, and likely create a code base that nobody wants to be responsible for.
• If using these tools in a team setting engage with them as a team. Share prompts, agree on a set of minimal instructions. Make sure everybody understands context and default prompts. Set guidelines and expectations for how the team uses the tools and what quality markers are expected to be present and maintained.

There are hundreds of blog post out there about LLMs. If I had to sum up my thoughts: The latest LLM generation is impressive, and can transform how we work, but they can also change the level of agency that individuals exhibit. Use the tool, but don't turn off your brain, and recognize our bias to trust these tools.