Posts tagged: ai-development
21 posts
All tagsPutting my old heating system into Apple HomeKit
My heating is an older radio-controlled Honeywell system that HomeKit doesn't support. A radio stick I built and a controller on my server put it in the Apple Home app and Siri without a new heating system.
Hikvision cameras and my routers in HomeKit
A Homebridge plugin that plays the encrypted streams from a HiLook DVR in the Home app, plus a web console for the DVR and a private plugin that restarts my routers from Siri.
A team of AI agents getting classic games running on old Macs
How ten Claude Code agents, one of them a manager, and a plain checking script build, test and speed up the open-source engines for six classic games on old Macs, and what they have done so far.
Half-Life on old Macs
Half-Life 1 as one universal app for PowerPC, Intel and Apple Silicon Macs, built on the open-source Xash3D FWGS engine, with 25 mods installed by a second app.
More Markdown in ArtfulType
My fork of Action Retro's ArtfulType, a Markdown editor for 68k Macs. More Markdown in the styled view, printing, Find & Replace, and bootable disk images built on Linux.
KeeperFX on Apple Silicon
KeeperFX, the open-source Dungeon Keeper remake, building natively on Apple Silicon as an arm64 binary with the crashes fixed and saves out of the app folder.
What an architecture-review skill found in PeerTalk and BomberTalk
I ran Matt Pocock's improve-codebase-architecture skill over PeerTalk and BomberTalk. This post is about the changes I made based on its findings.
Optimising Quake II on seven old Macs
I forked yquake2 5.11 and used Claude Code to get a build tuned for old PowerPC and Intel Macs, the sister project to the QuakeSpasm one.
Optimising Quake on seven old Macs
I forked QuakeSpasm to see if I could use Claude Code to get a build tuned for old PowerPC and Intel Macs.
How I guardrail Claude Code for devops
Last week the Guardian ran a story about an AI coding agent that deleted a firm's production database in nine seconds. From the post-mortems, the only safeguards in place were project rules and the AI tool's built-in prompts. This post is about what I run on top of prompts when I let Claude Code drive devops on AWS: IAM, hooks, and a deterministic CLI.
Cookie is live
Cookie is a recipe app I built. It's now live at cookie.matthewdeaves.com for my own use, with a home hosting mode if you want to run your own copy. A short post about what it does and the work that went into making it ready for the 'sort of' open internet.
Three Levels of AI-Assisted Pentesting
A month spent pentesting two of my own projects with Claude. Three escalating levels: ad-hoc review, a curated suite wrapped in a harness, and HexStrike. Here is what worked, what did not, and how to set it up against your own repo.
My Virtual Pentest and DevOps Team
If you're building apps with AI and putting them live, you have to at least try to secure them. Here's how I did it with bash, open source tools, and Claude Code.
Pentesting with Rook
Anthropic added a companion to Claude Code called Buddy. Mine's an axolotl named Rook. I've been running a pentest suite against Cookie and its infrastructure over the past few evenings, and Rook kept catching things I was missing.
Guardrails for Claude Code: Hooks, Rules, and Constitutions
Claude Code hooks, rules, and a project constitution enforced via GitHub Spec Kit. A primer on each layer, with a real use case: keeping ES5 compatibility on a recipe manager that runs on iPads from 2012.
Rockport: A Self-Hosted LLM Proxy
I built a self-hosted LiteLLM proxy on EC2 behind a Cloudflare Tunnel. It routes Claude Code traffic to AWS Bedrock with per-key budgets, idle auto-stop, image generation, video generation, and image editing. The repo is public and I'd welcome feedback from anyone who works with this sort of infrastructure.
Building an Evaluation Harness with Claude Code
I built an evaluation harness to measure the quality of LLM-generated pixel art sprites. Five rounds of calibration, a stubborn intent-vs-reality gap, and some useful C tools for getting assets onto classic Macs.
I Ditched My Planning System for Fully Automated GitHub Spec Kit
I built 14 custom skills, 4 hooks, and 1.2 MB of phase plans for PeerTalk. Eight phases done, real hardware tested, proper results. Then I started again from scratch with Spec Kit across three repos. Automated builds, hardware testing on Classic Macs, and a cross-platform chat app built in hours.
Automating Performance Tests on a Performa 6200
How MCP servers and Claude Code skills turned manual hardware testing into an automated workflow. From FileZilla hell to one-command benchmarks on a 30-year-old Mac.
Project-Specific Claude Code Customisations
Building custom MCP servers, skills, hooks, and agents for Claude Code. A workshop with Anthropic led me to create a full development toolkit for the PeerTalk SDK project.
Rapid Prototyping with Claude Code: A 3-Day Challenge
A 3-day challenge to build a working prototype with as many features as possible, plus a day of CI/CD automation. Django backend, React frontend, ES5 interface for legacy browsers, and AI-powered features.