Field Notes

Inspiration

I have been building projects using AI for the last year and a half or so, as well as using AI to help me on schoolwork. It has been a huge unlock both in the amount of work that I can get done in a given period of time, as well as strengthening the mental model in my mind of where everything lives.

While AI is incredibly helpful, it does also create a level of doubt in my mind about how much of the work and output that I have created is actually being retained in my head as knowledge. I think because of the power of the models, it can be really easy to abstract away a lot of the core learning that you get done while working, which isn't ideal, especially when trying to build skills as an early engineer or a student. Especially on the last point of students using AI to do their homework or help them on other tasks, this can become extremely dangerous if students give too much of their thinking and output to the models instead of doing that themselves. It could create an age of students who go into the workforce / whatever their post-grad opportunities are and don't have the skills that they're supposed to or that they claim to.

Field notes presents a solution to that problem by being the support network for students in demystifying some of the work that is being done by their AI models. It also extracts specific learnings for different subjects, abilities, and skills that students can tailor to themselves.

The two main tranches of learnings that field notes gives all users are:

• CI/CD and git-based best practices and learning
• Best practices for software engineering and agentic engineering ## What it does Field Notes triages through all of the user's conversations with AI agents and their GitHub commit history. It specifically searches for learnings that may be lost in the flow of building a project. ## How we built it Field Notes was built using GPT-5.6 models as well as Claude Code in an agentic building sprint. ## Challenges we ran into I ran into a challenge of deciding what the best model for Field Notes would be. Because this triage could be done many times and it's done autonomously, I wanted to try to preserve the amount of credits/tokens that were being applied to the product. Therefore, my first intuition was to use the cheapest model possible, which was Claude Haikou 4.5, but it lacked the reasoning needed to handle a task like this in terms of being smart in creating learnings. This resulted in a bloated database that had duplicated learnings and low-quality insights.

To amend this, I switch to Sonnet 5, as it offers much better reasoning and is only around double the cost, which sounds bad, but the cost is already really low. A Field Notes run on Haikou is around 12 cents, but a Field Notes run on Sonnet is around 22 cents, so it isn't that bad.

Accomplishments that we're proud of

I'm really proud of the Vercel deployment, especially because I wanted to make sure that it maintained the ethos of Field Notes: to never upload all of your agent context and learnings to model providers or into a space that you didn't want. I had Codex help me make a separate runtime for the judges to use so that there's data privacy for anyone who wants to run the demo.

What we learned

I learned a lot and am still learning in this implementation of Field Notes. I had to navigate creating adapters for many different coding agent sources, dealing with how to extract specified learnings from all of those sources, as well as normalizing them into a source-agnostic standard format. This showed me how even just data ingestion can be a roadblock to a project, as the codex and Claude threads were really simple and had a documented way to extract them, while Grok threads currently have no structured path. Not getting bogged down by something like this was a really good exercise, as it's completely based in reality - everything won't be accessible when working in nascent spaces.

I spoke with around 10 of my friends who are about to start / have just started different FDE, SWE, and other full-time technical jobs, as well as two established engineers at Google (L5 and L6), to get more context into how Field Notes could fit into their workflows, as well as how they feel their learning has changed since using AI agents for coding. These discussions showed me that Field Notes actually does have valid use, as they were all pretty overwhelmingly excited by the opportunity to have a tool like this. It also showed me how disparate the types of concerns, issues, and desired use-cases that a product can have even within the same category of users as the value that some of my friends thought that they would get from Field Notes were different than what the L6 engineer at Google wanted, for example. This was like a mini version of feature ideation and user feedback, and being able to do that with a personal project was also another really good muscle to train.

I learned about how to work with MCPs and even hosted my first MCP server as part of the extraction process for Field Notes learnings. That was really valuable because I think terminal and CLI tools are only going to continue to proliferate as the use of AI agents for coding and other knowledge work tasks increases. This process showed me how simple it can be to empower agents with products that I build in the future. I also adapted some of the search process from Claude-Mem, which is a developer tool that I have used for a few months, and it was really fun to mirror and extend the functionality of something that I already use.

What's next for Field Notes

I want to really improve on the educational features and usability of Field Notes. Right now, I think the core functionality is good and already helpful, but it still lacks some features like being able to connect to Canvas, or authenticate with a student email, or creating lesson plans based of Field Notes, that would make it a really easy surface for students to interact with. I'm going to build it out to be the true support system for young developers and students that I was talking about, and have it ready for a launch at the start of the school year in August.

Built With

• claude
• codex
• mcp
• python
• typescript