6 October 2026

A RAG chat assistant fully on Cloudflare (Workers AI, Vectorize, a Durable Object) where the contact offer has to be earned: the model only reports a judgment, code decides.

3 October 2026

I built a small agent that drafts replies to Instagram Reel comments with Claude Haiku and posts them only after approval. Writing it took nine pull requests. Connecting my account to the Instagram API, plus one extra account to test with, took longer: a disabled account, errors that said "not available" or "wrong password" when they meant "too many attempts", and two Meta app settings that turn every comment into an empty list.

30 September 2026

AI helped me build a SignalR update that was broadcast to every connected client instead of just the relevant one. The first fix it suggested, filtering in React, was cosmetic. The real fix was one line in the backend.

27 September 2026

Automating website deployment with AI: I tell Claude which site to add or update and what to change, and everything after that is automatic, except my review of the pull request. Claude gets no Cloudflare credentials; it prepares one PR with config, custom domain and images, and GitHub Actions and Cloudflare Workers Builds deploy it after the merge.

26 September 2026

My blog's API is an Azure Container App that scales to zero, with a 30 second cold start. How three layers of caching keep readers from ever waiting on it: a stale-while-revalidate cache in a Cloudflare Worker, a durable KV fallback for Cloudflare's per-colo cache, and Cosmos DB Change Feed pushes so reads stop waking the container at all. With real monthly costs, and the trade-offs I accepted instead of solving.

25 September 2026

Claude Code could change my frontend but never see it, so I was the one checking every layout fix in the browser. A search box that took four PRs in one day made the cost obvious. I added Playwright and headless Chromium to the container, so Claude can render the page and measure it before it tells me it's done.

25 September 2026

In my Claude Code container for new projects, the repos live on GitHub and there are no secrets in their history. So there, Claude creates its own feature branch off master and pushes it, and I review and approve the pull request. The setup is the GitHub CLI in the image, git using it for credentials, and a fine-grained token that lives outside the image.

25 September 2026

Claude Code works in a history-free copy of my brownfield repo, because the real repo still has old credentials in its history. Getting its commits back used to mean a temporary remote and a manual push every time. Now the real repo has a permanent, fetch-only remote pointing at Claude's copy: I fetch, branch off master and cherry-pick, and nothing ever flows back into the container.

23 September 2026

My last post covered copying a brownfield project into a clean repo and cherry-picking real commits back, dodging the secret rather than dealing with it. There's a more direct third option: rewrite history and remove the secret from the repo itself with git filter-repo. Here's how, what it costs, and why I still rotate the credential anyway.

19 September 2026

Replacing the default AI assistant image with my own avatar: training a Higgsfield Soul ID that stubbornly refused to stylize, switching tools to get a real cartoon render, and generating a small set of consistent poses for a crossfade idle animation instead of a video loop.