Claude-Code

Background Okay, so this post is a follow up to a few different posts, but most notably The Token Saving Cake is a Lie. Before reading this one, I highly recommend checking that out first. Go ahead. I’ll wait. Back? Good. Let’s continue. Headroom So I said that the solution is probably something like Lynkr, but before building the integration, I discovered Headroom (45k stars, Python, its own trained compression model, ships RTK as a component). It does everything the Lynkr plan described, supposedly better. Set ANTHROPIC_BASE_URL=http://localhost:8787, run headroom wrap claude, done. ...

Previous posts for context: The Future of AI Coding is Directing and C.A.S.T Progress: First Blood Last time I had one event flowing from a ticket board to a NATS topic. That was the plumbing. This time the plumbing has something running through it: a ticket goes in one end as a one-line wish, and a reviewed, merged, cleaned-up branch comes out the other. No human types a line of code in between. ...

Tools Full comparisons and context in my 2026 AI tech stack post. This is just what you need installed to follow the workflow below. Claude Code If you’re new, start with the cheat sheet and Anthropic best practices. Security — set this up first: Claude Code Security Hooks — 7-layer prompt injection defence, read guards, canary files Lock down your .env and any git-secret files in .claude/settings.local.json before anything else MCP: Context7 — library/API docs on demand DeepWiki — open source repo documentation Skills: ...

Previous post for context: The Future of AI Coding is Directing. Let Me Introduce You to Your C.A.S.T. Also, this banner was generated by ChatGPT and whatever model they are using at the moment, and I’m kinda impress, NGL. When I wrote about C.A.S.T last month, the design was solid but unproven. I had a SPEC, ADRs, hire/fire scripts, and a lot of conviction. What I didn’t have was a single event flowing from a ticket board to a NATS topic to a waiting agent. ...

Introduction In some ways, a lot has changed since my last post, and in other ways, not much has changed. There has been a big shift in the last 12 months, especially toward “agentic engineering” and how we run multi-agent AFK setups vs HITL as we have been. I touched on this in my introduction to the tool I’m working on as well, which, spoiler, is part of my toolchain. Kinda. It’s a WIP… ...

Just because your fancy new Claude Code plugin prints out “50% tokens saved” in the logs doesn’t mean you actually got them. I was playing around with Caveman Compression last week and tried to get the UserPromptSubmit hook to rewrite prompts using the tool for savings across the board. Turns out? The hook doesn’t allow it. Probably for security reasons; can you imagine the injection attacks if any random plugin could hook into and replace your prompts? ...

Introduction I came across caveman-compression — a tool that compresses text into a token-efficient “caveman” format before sending it to an LLM. The pitch is simple: It can cut 30-65% of tokens depending on the content, which at Claude Opus prices adds up fast. I immediately wanted to wire it into Claude Code so compression would happen automatically on every prompt. The obvious hook was UserPromptSubmit. ...

Previous post for context: Two AI Agents Walk Into a Codebase If you look at social media, everyone will tell you that most people are still using AI like a very fast keyboard. You write what you want, it writes the code, you review it, you move on. Faster, sure. But you’re still in the driver’s seat for every line, every file, every decision. They will also tell you that the future is not that. The future is swarms of AI agents doing the work of 100s of people. And I mostly agree. The shift that’s actually happening — and I say this as someone who’s spent the last month living inside it — is from writing code to directing agents. ...

That’s the punchline and the problem. I’ve run into these two similar scenarios which have had me stumped and annoyed. In one, I’m running Claude Code in both GoLand and PyCharm, trying to integrate one project’s API with another project’s GRPC service. In the other, I have a Scala library that needs updates, and then multiple services that depend on that library. The hiccup: each instance is completely unaware of the other’s existence. ...

Intro I posted over on LinkedIn about an interesting comparison between Lovable and Bolt.new which I did. I had this idea for my next Flutter project, but after being hit by the 5th LinkedIn post about Lovable in as many minutes, I had this thought. 😂😂😂 No, actually, the thought was more along the lines of ...