The best place to show the model the context it’ll need next time is right next to the code. So I’m writing long block comments now. I’m finally ready to work on the Linux kernel...

Every PR I review is a mix of human and AI output. But often it’s impossible to tell where human or AI intelligence was used in the process.

The software engineering industry tends to change its norms relatively slowly, usually only after doing things the wrong way for a while and feeling the long-term pain. AI has changed nearly everything in just a year. But the norms will take a while to catch up.

The biggest challenge with AI coding is likely to be: building software is still a team sport. Poor use of AI by some team members pushes the real work onto others, who get “splashed” by slop.

I set up a cron job to send me the GitHub notifications I actually want to see. Took 20 minutes. Genuinely baffling that GitHub still hasn’t implemented this.

Ghostty fixed emoji input on macOS just one day after I raised an issue with a prototype fix ⚡️ Emoji inputs suck on macOS, seems likely I’m one of only a handful of people who’ve bothered to set it up! github.com/ghostty-org/...

I’ve been experimenting with using emojis and short acronyms in my `AGENTS.md` file: 🟢 = Babysit the GitHub checks cip = Commit and push 🧵 = Do this in a tmux window Works well

I’m a big fan of the Pi coding agent, but it needs some extensions to get the most out of it. Here’s my setup: github.com/richardgill/...

I spend a lot of time reviewing AI-generated code, so I took a few days to improve CodeDiff.nvim, my favourite diff tool for Neovim. I added GitHub-style diffs, explorer line stats, and fixed 10 bugs and performance issues. My fork: github.com/richardgill/...

Just finished building a 3d printed keyboard. Wrote about it here: richardgill.org/blog/3d-prin...

Trying out pi as a replacement for claude code. It has extensions, so you can really customize it. Here's 'Reveal' by @mitsuhiko.at which finds files mentioned during the conversation and opens a custom ui to let you open files or add them to the prompt. buildwithpi.ai github.com/mitsuhiko/ag...

Making exam cheat sheets with Claude Code: • Converts textbook PDFs → Markdown • Suggests Typst for clean layouts • Ingests textbooks + past papers • Iterates in a feedback loop • Exports a gist for voice Q&A in @ClaudeAI It’s such an amazingly general tool!

Claude

Talk with Claude, an AI assistant from Anthropic

claude.ai

Claude Code can solve Advent of Code 2025 in 2 hours - no human intervention required. I wrote a post about why it's so good at 'LeetCode' (includes a 2 hour video of Claude solving it): richardgill.org/blog/solving... @anthropic.com

There aren't enough easter eggs / quirky things in software these days but @anthropic.com has some good ones: Claude Code coloring in 'ultrathink' is fun! /stickers lets you buy stickers, and discover the mascot is called 'Clawd' www.stickermule.com/claudecode

Was updating my Claude Code rules and remembered I'd written this guide for my junior engineers when I was running z-dev. github.com/richardgill/... It's been 8 years, but most of the advice is timeless.

Just released 0.0.1 of Patchy. It's a CLI to help manage patches. I'm not a fan of long-running Github forks, I use Patchy to maintain my fork of Firefox. github.com/richardgill/...

GitHub - richardgill/patchy: Command-line tool for managing git patch workflows

Command-line tool for managing git patch workflows - richardgill/patchy

github.com

claude "/stats" is the new GitHub commit heatmap

Wrote up some tips for using Neovim with Claude Code.

Xata - Postgres at Scale@xata.io · 11mo ago

Using Neovim with Claude Code? Xata engineer @richardgill.org shares a practical setup to: - See agent-written changes instantly - Reload git diffs in real time - Yank code with file locations so Claude gets proper context 📝 Tips for configuring Neovim for Claude Code xata.io/blog/configu...

Took the plunge to learn Nix and configure a mini-pc with it. github.com/richardgill/... Definitely learned a thing or two about Linux along the way!

LLMs are currently kind of slow. OpenAI's new gpt-oss 120b is a pretty capable model in the benchmarks (maybe ~1 year behind the state of the art) If you run it on Cerebras you can glimpse the speed improvements that are coming.

Dev tooling often behaves differently if you set CI=true We're going to need AI=true so that dev tools can behave differently when called by an AI agent: - Non-interactive (like CI=true) - Less logging to preserve context - Tell the agent what commands it could run next based on what happened

When you use claude code a lot, it feels obvious that coding with agents on mobile makes sense. Seems likely >20% of 'coding' will happen on mobile within 5 years (presumably from ~0% today!)

If you have a social media app where the DMs are important but you want to stop being exposed to the feed. e.g. LinkedIn or Twitter You can use beeper.com which aggregates chats across multiple platforms. Then block/uninstall the original app.

Was explaining agentic coding to my mum yesterday and found myself saying: "My job used to be typey typey, and now it's chatty chatty" Kind of sums it up nicely.

With Claude Code you can bring it up to speed on a PR by doing this incantation: git fetch origin refs/heads/main:refs/remotes/origin/main && git diff "$(git merge-base --fork-point origin/main HEAD)" Which calculates the PR diff to main, like you'd see on Github.

Freshened up limitphone.com (again!) ✅ Visual refresh ✅ Landing pages for parents, gambling and porn ✅ New main video + videos for parents, gambling and porn. ✅ Positioning v5!

Using Claude Code on the Max plan feels a bit like fast-forwarding 2 years in the software engineering industry. It's equal parts exciting and scary.

LLMs do markdown code blocks way better if you ask them to put it in quadruple backticks.

Something I've learned about myself in the past few years is to trust my 'spidy senses' on complex problems. They turn out to be right 95%+ of the time but come with 0 rational explanation I can pass on to others. Usually takes a couple days for my brain to catch up.

I'm internally conflicted about using AI tools to code. I want it to be good! But I keep trying claude code in good faith but remain pretty unconvinced. I'm slowly diverging from quite a few of my friends and colleagues who seem are all increasing vibing.