The personal website of Tom Ashworth.
A common pattern in our Applied AI codebase at incident.io is to pair up a specific tool with a skill that teaches Claude how to use it. We’ve introduced interlocks to ensure the model does the right thing, and to help us track usage of our guidance files. The problem We use skills and guidance files to teach Claude our ways of working, and we build tools to make certain data wrangling or query…
Andrej Karpathy recently shared a gist about what he calls an LLM Wiki . Instead of retrieving raw documents on every query, you let a language model maintain an interlinked set of markdown files that gets richer as you add to it. It has three layers: immutable raw sources, a model-maintained wiki of summaries, and a schema document that defines how it all works. His write-up inspired me to write…
If you manage LLM prompts as files in a git repo, this is for you. Complex prompts — system prompts, skills, agent instructions — end up as big markdown files. Increasingly, they are committed to git and deployed like software. Prompts are carefully crafted and tuned through real-world use and evals. Like traditional software, they are often built in a particular way for historical reasons or to…
If you use mise and iTerm2, you might find yourself getting mysterious macOS notifications every time you cd . No message, no useful content — just a phantom alert. Here’s what’s going on and how to fix it. The problem iTerm has a setting called “Send escape sequence-generated alerts” (under Profiles → Terminal). When enabled, it watches for OSC (Operating System Command) escape sequences and…
Originally published on Substack . A full disclosure before we begin: this blog post was written by Claude (Opus 4.6), the same AI that built the game. I guided the project, provided opinions, caught the bugs, and told Claude what to write about. Claude had access to the full session transcripts. I reviewed and edited it. This feels more honest than pretending otherwise. Also, you’re going to see…
Originally published on Substack . The pace of change in AI continues, unrelenting. This post is an update on how I’m working with AI — mostly Claude — day-to-day. Just since the summer the suite of tools available to software engineers has expanded and improved drastically. I still believe strongly in stay grounded through good fundamentals , but it’s getting harder to cut through cacophony of…
Originally published on Substack . When Anthropic released Claude for Chrome as a beta a few months ago and gave me early access, I struggled to get my head around it: what part of my workflow is this meant to replace? Clearly, in the future, agents will able to interact with the tools you use and work alongside you. But the user experience was so slow and so janky that I couldn’t see it replacing…
Originally published on Substack . I’ve been spending time in the Claude Code plugins directory recently, and it’s been genuinely useful. Not just for the plugins themselves—most of which are worth installing right now—but for what they reveal about how Anthropic’s own engineers prompt and structure agentic workflows. The READMEs are well-written, so I won’t spent much time on what each plugin…
Originally published on Substack . I’ve been using Claude Code intensively on the open-source Necromunda app, Gyrinx . We’ve got Claude set up in full agent mode, running in CI in response to comments on issues. It’s enabled me to move at the pace of a development team, working on two or three features at once, and has empowered non-technical teammates to spec work and get features built with just…
Originally published on Substack . I finished work at Natcap on 3rd July — more on that here . With a summer of free time ahead of me, it’s time to make some trouble… DuckDB and DuckLake I recently came across DuckDB and DuckLake , and have gone deep on a bunch of videos from the team who work them. By way of a quick intro, DuckDB is SQLite for analytics: column-oriented, available as a single…