If you rely on global Claude skills on your local machine, you need to make sure that they don't irreversibly change in behavior. If your LLM creates and maintains your skills, there is a chance this will happen. One useful thing I started doing is versioning my global Claude skills, such /grill-me, /notify-slack, /write-a-personal-skill, /pr-review-difftales, etc. Usually, global Claude skills live in the ~/.claude/skills folder. It's Claude's house, so I don't want to go in to version it with git. I think you will just be fighting Claude and .gitignores and includes. Instead, I have another ~/code/github/dotclaude folder versioned in git where all the skills and other cool "local agent harness" features live. My skill-writing skills know to symlink them to ~/.claude/skills so they are available in Claude chat instances via the / invocation, so it's fully compatible with my regular Claude Code setup.

AI Content Nuggets Week#25

SaaStr AI - How We Built a $10M+ AI Agent Stack (614 Meetings Automated) - I follow every podcast of SaaStr.ai for a few months now. They have a high ticket event business that they run with "3 humans a dog". This episode is a great intro into their world.

Todd Dickerson (@ToddDickerson) - The best thing I did for my AI setup was eliminating agents - Ditch the swarm (the chaos in my case :D), run one super-agent with deep context.

Wasn't enough to bother you with an email, but last week I rambled a bit about MCP design content a little bit if you need this for any reason: link.

LLM evals

I'm doing my home-baked evals when building bigger features, i.e., letting the agent test the feature for some number of rounds (depending on the feature's complexity), evaluating what went well and what didn't, and improving based on what it learned, including a report. So I'm looking to tweak my approach to make it more effective, token-efficient, and deterministic. But evals are useful in different places, so thus far I've only found a talk relevant to LLM model builders, and the other was a product demo.

AI Engineer / Nicholas Kang & Michael Aaron (Google DeepMind) - Agentic Evaluations at Scale, For Everybody - Evals for model builders, not as relevant for product builders.

AI Engineer conf / Dat Ngo (Arize) LLM Observability, Evaluation, Experimentation Platform - just a product demo, but seems like a funky product giving you insights (including "traces") into the life and movements of your AI system.

Just for fun
(I'm having fun so you don't have to, you can just take a quick look at the wondrous yellowy links)

Mo Bitar - Chinese OPEN SOURCE model better than FABLE - My recurring AI-hype detox. This one was a bit balanced, and Mo starts trying some stand-up in the first couple of minutes. Full support here despite the anti-AI stance. 💪

Alberta Tech (@albertatech) & @Mytechceo - CEO finds out tokens cost money (https://www.youtube.com/shorts/9wwbjodvB9A)