viewing probably work posts

Decided to shut down Limoncello. Mostly because I I agree with this blog post about plan mode being dead — the agents are good enough to just execute.

I still think there is an opportunity for an agentic accessible backlog of ideas though.↳

You know a frontier lab is becoming serious about robotics when one of them acquires Matic↳

Era: a Personal Finance MCP ↳

I really like this MCP — it basically acts as middleware between LLMs and Plaid, allowing you to connect your bank accounts to Era, and then have LLMs connect to Era via MCP to access those bank accounts.

Now, sure, there’s a lot of risks to doing this in terms of giving LLMs access to your financial data, but if you did want to give Claude your financials, this is the way to do it!

The service also has a lot of pretty clever functionality such as categorizing transactions that then gets applied across whatever LLMs you have connected.

✰ Markupdown

A simple tool to leave comments on Markdown text for LLMs

Something i have been thinking about: the logical entryway for AI for consumers is on the iPhone, not an app but built into the OS — or maybe, built into iMessage. It would be super great to be able to chat with (AI) models and harnesses alongside my other iMessage chats. The key difference? it would be effortless and where people already spend their time. Maybe that’s how agentic B2C AI blows up.

And what’s the equivalent for B2B? It’s Slack.↳

Projects in Github Personal Access Tokens ↳

An unfortunate bug with GitHub is that when creating a personal access token you can specify granular permissions for just about anything except the ability to create and manage GitHub Projects.

The only way to create a Personal Access Token for GitHub projects is to move the repo into its own organisation, and then you have the ability to specify these permissions.

I find this a truly bizarre omission to the point that I assume its a bug: I can create a PAT for issues but not projects?

Hopefully this gets resolved, although I have been asking for this for all of 2026.

How to Spot AI Writing ↳

First, consider words. The vocabulary that bots overuse has changed: they no longer “delve” and there are not as many “tapestries”. Instead they offer a significant number of polysyllables like “significant”, “increasingly” and “consequences”. They use more rare words (“interdependence”, “reindustrialisation”) and scientific lingo (“parameter”, “methodology”) than humans, and are fond of nominalisations (making nouns from verbs, such as “expansion” from “expand”). All the llms in our study use such words, but particularly Gemini and Claude.