Agentic LLM Skills Help Solidify Notes into Ethos

Note-taking has always been a key part of how I learn.

I used to have a “the more the better” approach to my notes, but the more volume there is, the less easy it is to find what I need when I need it.

My dotfiles repo has become my de facto knowledge base for the last few years. My engineering notes are taken largely in the form of code these days:

functions
aliases
scripts
configuration files

If I do keep a plaintext or markdown file, I usually add some kind of tooling on top of it to readily fzf through its lines (if applicable, otherwise ripgrep finds things just fine).

The loose end for me for in this system has been less-deterministic, more philosophical notes. Checklists, heuristics, and other soft knowledge that doesn’t fit neatly into a code file.

Since I’ve started to develop custom agentic skills, I’ve found this fuzzy area of knowledge retention suddenly has a home. I can translate my notes into LLM skills, which has a number of benefits for me:

Concision

Since token count matters, I am forced to distill my notes down into the things that matter most to me. I must trim the fat and focus on the core ideas.

Not just useful for AI

I review these skills regularly (made easier by the forced concision), and the act of reviewing them helps me solidify my knowledge and fine-tune skills/notes over time. Yes, it’s useful to have them invocable with AI, but I’d still want these notes handy even if I didn’t have an LLM to work with.

Clarity and Retention

I use skill creation to work through my thinking on topics that matter to me. The act of skill creation is the act of decision-making and opinion-formation. If I learn a lesson when building a project, reading a post, or listening to a conference talk, taking note of it in a skill helps solidify it and ensures it’s not just transient knowledge.

Forcing Concision Results in an Ethos

Noting that I lose searchability when my notes are merely additive and not distilled and regularly reviewed, the things that matter most to me become apparent from my agentic skill definitions.

To be explicit about it, this ethos includes:

Building tooling for the tools: so much of what I build is to make the next thing I build more quickly, reliably, securely, and with more fun involved.
Doing things the “correct” way: “working” isn’t good enough, I want things to be clean.
Being in control of the specifics: I want to know what, why, and where things are happening in my systems. LLMs augment my work, they don’t replace it.
Ensuring process discipline and security hygiene always.
Maintaining a product sense, not just a code sense.

I’ve been somewhat surprised to discover just how useful creating a repo of personal agentic skills to has been beyond just invoking them with an LLM. Definitely give it a try if you have the same sensibilities about notes and recoverability I’ve shared here.

Song On Right Now: “190304-08” by Hideyuki Hashimoto