unknown

Fantastic write-up on SKILLS by the GOAT @trq212

I have been leaning into SKILLS a ton too, both for my Claude Code setup, and also for building agentic software (mostly with DeepAgents from LangChain). 

You should read his post top to bottom, and if you don't have much time, skip my comments and just jump straight into it. Here's my stream of conciousness on them. Note that I use them mostly to orchestrate agentic software and to do compound engineering.

- I use them a lot for things like code review, session logging, auditing my documentation, distilling information from other sources (e.g. adjacent repos), and updating my stored context automatically

- I rely *heavily* on progressive disclosure. LLMs like a 'map of the territory' and will happily navigate your files. Use this to your advantage within skills. Any single MD file longer than 100 lines is a smell to me.

- Use skills to clean up and improve your skills. For example, my /update-docs skill makes sure that other skills are up to date with the current info in the repo, and actually do progressive disclosure

- Using hooks to run skills, and relying on CI/CD to run skills, can make the feebdack loop / management of your context much better.

- You can distill skills from your claude context directly. Claude Code stores all of its sessions and plans. You can have it introspect itself to help you come up with good skills to build.

- Understand the front-matter because it is important (e.g what tools a skill can use, how it gets executed - e.g. fork)

- Skills themselves should use progressive disclosure internally. Map of the territory all the way down.

- Bake feedback loops into your skills. Most of my skills have some aspect of generating distilled context that my agents can use later (e.g. adding to a WORKLOG.md, or storing a distilled version of a session in a session_{xyz}.md)

- Don't over-prescribe stuff. My early SKILLS were *way* too long, and now models are good at just figuring stuff out. Let them take the wheel a bit more

- If you are using LLMs as part of a product (e.g. not just Claude Code, but you use something like Claude Agents SDK or DeepAgents) - write your traces to the file system, and build SKILLs that inspect those traces - this is my primary way to make my context better. Basically an automated version of hill-climbing

- If you ever find yourself trying to bootstrap a thing yet again, figure out how you can turn that into a SKILL

I know this was super disorganized and a bit rambling but trying to timebox this to 5 minutes because I have other stuff to do. Probably missed a lot and also still have lots to learn. Either way, do skills stuff. It works great.

A retweet praising a write-up on SKILLS, mentioning its use in Claude Code setups. Minimal content beyond the endorsement.