Pretty sure Claude just saved us an additional ~$1,000 in taxes. (I used Cowork to review the draft return our tax preparer prepared.) We really are in a new world when it comes to the informational costs of policy compliance.

It seems really clear to me that modern agentic AI software development tools like Claude Code mean taking a government legacy system and making it easier to change is much, much cheaper (and more feasible) than before. Who is doing this now?

I think one of the more interesting questions around AI right now is what types/shapes of new organizations it enables. Who's doing the best thinking-out-loud on this?

The best quality of life use of AI I have now is a little tool where I can just go for a walk-and-think an use Siri to save thoughts and ideas, and then have Claude process them based on type (eg this is research, kick off a research task! this is an idea, riff or propose an MVP!)

A piece of advice that I strongly believe in public/civic entrepreneurship is that it can be very effective to dial in on one key issue that basically no one else is working on and go hard at it. I think the first folks who apply AI leverage to this will have massive public positive impact.

People who work on government technology need to be taking much more seriously what it means if AI enables a small team of permanent staff to fully own frequent changes to the system, rather than managing vendor change requests.

A salient point made a decade ago in civic tech thought was that people's expectations have been elevated by consumer tech relative to government services. Worth wrestling with what an even bigger gap—including with agentic volume increases—could similarly imply.

I was learning a bit about pandemic preparedness and have discovered a new vector to build defenses against: The SF Cal Academy of Science on a day that daycare and schools are closed.

@econberger.bsky.social do you have any go-to for empirical scenarios of AI caused job losses? The impetus for my question is starting to build out what the scenarios would imply for benefits programs (eg SNAP, UI) and what one would expect the second order impacts of that to be as a result

We've had user research participants telling us how they use LLMs to summarise their NHS App health records for them.

People are wondering when AI will be helping people navigate government services. The answer is, uh, right now. Here's Gemini (built into Chrome!) giving both accurate and helpful information to a common and high stakes question on a SNAP application.

I really don’t like snark and try to minimize it in my life these days. But… Google’s model doing bad things it’s aware are bad in pursuit of maxing out KPIs is… pretty darn funny to me. (Also this paper scares me, as does much research in this area honestly.)

Fixing government technology outcomes is not possible if one rejects the very premise that it is a market, and you need market interventions on both the demand and supply sides.

A bit of tacit knowledge on a Friday: it's long been a known heuristic that if you're doing innovation/tech change stuff inside government, often the political leadership wants something flashy but not so deeply impactful - a flash tax. Take note: AI is very good at generating flashy web apps.

If AI capabilities are really on the exponential, then the constraint for using them for true public goods in the economic sense becomes giving the government the agency (capacity) to actually do things. State capacity = capability + agency

If you have a persistent AI agent with memory, I highly recommend assigning it to read James C. Scott's classic, "Seeing Like A State"

There’s a surprising amount of criticism/FUD around government use of AI that, when read carefully, rounds to the assertion that government cannot be trusted to do effective management at all.

I built a little web app that makes it simple to send a photo postcard to friends/family (and the outbound card has a QR code that makes it easy for them to reply with their own photo postcard back) Such a delight. Seriously, use Claude Code to build stuff you want to exist in the world.

The most *practical* value I've gotten out of Claude Code is writing a bunch of helper scripts for Asana (where I manage my life + our household stuff). It's mostly grinding out friction, but for tasks I do hundreds of times a day... awesome. (And I usually run on Claude iOS app; no server needed)

A small-but-mighty trick for using LLMs for cooking: - Use an AI voice transcriber app to just walk through your pantry and capture it all by just talking - Do the same for fresh fridge stuff - Now you can give the LLM *way* more context for what to cook tonight

File under things that are wild to me: Claude Code can't access congress.gov pages because their Cloudflare config blocks all bot traffic.

Pretty sure today marks the first time a current or former Medicaid Director has ever used Claude Code. (Was pretty darn cool to facilitate it.)

I think those who want to do serious work on fixing government procurement outcomes should be most focused on what on the government side makes rational optimizing vendors do what they do.

One of the more interesting things to me about AI agents is they potentially open up a market structure where a payer pays for an outcome, rather than software (or change requests on software.) For AI transition philanthropy this could be interesting: market making for current market failures.

Candidly I can't believe Google continues to fumble its user advantage on AI so badly. Gemini in Google Docs... can't directly change formatting in the doc? (This is could be an interesting eval to track just how far behind Google's inertia is putting it.)

I'm early in the exploratory phase, but it's so interesting to give Claude Code a SNAP automation project scope that is illegal under federal rules (but not implausible) and see it refuse to build it. (Interestingly, if asked to move barely >30 day apps to the end of the queue, it accepts.)

A Friday thought: I reject a definitional shift of civic tech into something more closely resembling modest organizational change management.

My partner was hospitalized and I now have the strong desire to build an LLM eval for guidance on avoiding surprise bills from ERs/hospitals/ambulances. High quality advice in that (charged) moment would be deeply impactful at scale.

Great example of a tiny life improvement AI coding (Claude Code) made a 5 minute project: Update my Siri-SMS-to-Asana tool to put any message with "today" in it in the Today section of My tasks

A prompt I gave a friend that he rated highly: “What would you build assuming software development is free?”

I predict one of the stranger aspects of building software becoming free is that a key constraint will be a person’s own individual awareness of what they really want (and honesty about their intents)

"Claude says USAA" A text from my partner just now. And also a concise signpost of where we're headed.

A note I jotted down today: “Don’t try to persuade the skeptics; try to intrigue the curious”

Comments from software engineers diminishing vibe coding strike me as a signal of lack of understanding of the true spectrum of software: from super powered Excel spreadsheets to systems defending against state advanced persistent threats. It’s far, far from a binary around an imagined “real”.

The first AI lab to release a "circling back" capability will decimate the software PM market.

Okay I do think my favorite use of a language model at this point is building a pipeline for Claude to ingest a recipe photo from a cookbook and unpack all the tacit cooking knowledge into explicit knowledge I can actually act on!

I built a Skill to access SNAP's federal regulations in AI agents as part of some bigger work. It was really enjoyable to see how easy this has become. (I'll confess to getting a bit turned around in the Claude docs on skills vs. commands vs. plugins)

I've been wanting to try the Claude Agent SDK more directly and ended up building a tool that ingests recipes (from cookbook photo) and does a bunch of things, but the best one being making implicit aspects of the recipe explicit.

A state capacity + AI thought: 1. Much of the dysfunction attributed to governmental organizations is more accurately attributable to *large* organizations generally 2. If AI can give outsized leverage to small teams, what could that mean for effective government?

Something I've been taking scratch notes on recently: I think for diffusion of AI capabilities to reach people accessing government services/benefits to actually happen quickly + go well, we might want some "responsible agent standard" specific to that domain's nuances. (I have spec thoughts!)

Using Claude Code I built a tiny little Apple Watch app to talk to… Claude! (via Siri) It’s called Norbert (cybernetics!) and has been super helpful for on the go tiny quick questions and staying off my phone. Getting it in TestFlight, DM me your email if you want to try it!

Having a toddler, a relative with early stage Alzheimer's, and a senior dog who is diabetic and blind, I really genuinely think if AI takes away a lot of the "solve small puzzles on a computer all day" we'll have quite meaningful work to do.

Well the toddle has pneumonia and I’m taking the dog to the pet ER. Happy new year and may the exogenous shocks of this coming year be fewer than in the prior!

I’ve now got Claude Code in the iPhone app running Python scripts under the hood to access our household management stuff in Asana and help me go through stuff on the go.

I’m updating my view in working in the open and open source. Both seeing more second order effects and the drastic reduction in cost to repurpose open code/work from AI has made it much, much more valuable in my eyes. open.substack.com/pub/daveguar...

AI helped me successfully transfer my toddler. I heretofore fully reject any claim that AI is useless.

I am so, so excited for a world where the creation of software is not constrained by an individual’s faculty with language syntax and tolerance of yak shaving.

Yet another helpful use for Claude Code: connect it to our household tracker (Asana) and go through open items and propose what an actionable, concrete next step is.

Civic tech could stand some more consequentialism.

I am SO excited to learn that Sefaria has an MCP server! This serves exactly the niche need that I and 7 other people have!