Dave Guarino

Yet another use of Claude Code I probably won’t go back from:

Working with an API I don’t know, just tell it to comprehensively read the docs and come up with some options for me

(Solved a very annoying subtle bad default for me last week)

We need to start talking about an AI benefit design And more specifically: how we design it to actually work

We need to start talking about what an AI jobs displacement benefit program would actually look like.

Yeah, I mean, they're all shape-of-system things at the end of the day. Just have to spend my waking moments translating for the less gonzo crowd I suppose.

An essay I'm simmering on — thinking about government's use of tech (& AI), disambiguating problems of:

- Capability
- Alignment
- Dysregulation (cc @dsquareddigest.bsky.social)

An important differential diagnosis for interventions! And one that I think is implicit (and/or wrong) too often.

AI Models Are Getting Dramatically Better at Complex Policy Questions: Evidence from SNAP Asset Limits Testing 45 AI models on SNAP eligibility questions reveals dramatic improvements. See how GPT, Claude, and Gemini evolved from harmful to helpful answers.

New research post! We tested ~45 AI models released over the past 18 months on a hard SNAP policy question (one with a meaningful potential harm profile.) We saw *dramatic* improvements in how they handled policy complexity.

www.propel.app/insights/how...

One meta thing with governance of gov use of AI I feel the traditional policy world is ignoring a bit is that the AI capabilities at the frontier actually can be a huge tool in doing oversight, monitoring, etc. scaling accountability

Kernels 02: Claude Compadre (Claude Code for non-coding things) Also: is curiosity all you need?

Kernels 02: Claude Compadre (Claude Code for non-coding things)
open.substack.com/pub/daveguar...

Having my newsletter emails say "Dave Guarino from Dave Guarino's Occasional Newsletter" has annoyed me long enough.

So I hereby announce my newsletter's new name:

"Plausible Legibility"

More progress on using Claude Code for my absolutely-personally-tailored daily todo list tool.

First bit was lots of keyboard shortcuts a la vim. Next I want a voice interface (via Apple Watch) so I've prototyped that with a text-based chat interface that will eventually be driven by voice.

I've been picking at an idea, that a corollary to "idea machines" in certain problem spaces is the creation of "agency machines"

In domains with high stuckedness, sometimes you need to inject a new actor with lots of ability (and incentive) to do things

I did it.

I waited out the need to actually learn the Javascript ecosystem.

Enjoying this AI coding future EVEN MORE upon this realization.

(lol I pay for Max...)

wrote my own static site LLM client in a few hours that does exactly what I want and nothing more. gpt-5, claude sonnet and opus, gemini 2.5 pro + web search + reasoning toggle. it rules. will make repo public soon

so good

I'm using Claude Code to add vim-style keyboard interactions to a personal todo app because THE COST OF PERSONALLY TAILORED SOFTWARE HAS BECOME $20 PER MONTH.

I’m beginning to suspect one of the highest societal value things to do right now given AI is to make the work displacement effects it will have (is having…) legible and indisputable.

I will go out on a limb and say that I do believe we should use new AI capabilities to increase the capacity of the state to deliver on the goals democratic governance processes have arrived at.

I'm increasingly convinced Claude Code is an early glimpse of what general purpose AI with agency feels like.

I'm experimenting with making it usable for non-programmers:

- Created a custom output style called "non-technical user"
- It explains what it's doing in lay terms

I went to Berkeley so long ago that in my day effective altruism was just called “applied econometrics”

This is very good to hear!

Started reading the Expanse series (Leviathan Wakes) and really enjoying it, particularly the political economy dimensions!

I'm going to write up my Claude Compadre™ thing (using Claude Code for non-coding things) but a few tips:
1. Have it take notes in local Markdown files, and self-update Claude.md based on what it learns
2. Use Opus 4.1 for the initial planning (can pay per token by API just for this with /login)

Me: "Man, these Claude usage limits can be a bummer! Opus is so much better!"

Also me: "Write a satirical dialogue in the style of Armando Ianucci about Krushchev's visit to a US supermarket. Think ultrahard."

I've been using a pattern I call "Claude Compadre" — using Claude Code, but for decidedly non-coding tasks.

This approach means I can have Claude quickly "grow" its own capabilities via code generation.

Example here: generating code to hit the Asana API to help me with family task planning

Somewhat strangely, I'm finding the Asana MCP server does not work well, whereas when I just use the model to generate code that directly hits the Asana API it works fine. Seems like a design/implementation gaffe?

It is time. Hello, Claude, my friend.

Claude Opus 4.1, to me, while building out a comprehensive policy understanding of SNAP simplified reporting:

“The confusion in the documents shows how complex this system is - if we're getting tripped up just researching it, imagine how confusing it must be for actual SNAP recipients!“

EXACTLY

Using Claude Code to check logs for suspicious activity

I think using AI as a cooking assistant would get a lot more people seeing the value these odd alien intelligences present in their current form.

Final judgment on Piranesi: second half much more enjoyable, but, to me, the reveal was not worth the slog.