Writing
Architecture write-ups, patterns and cost math from wiring Claude into apps that were already live.
You have a Laravel app shipping revenue and a roadmap that suddenly says "add AI". Here are the 5 highest-ROI insertion points — with patterns, code, and the failure modes I have actually hit in production.
A walkthrough of the AI-agent system I built to run my own job search and freelance pipeline — five scrapers, a Claude scorer with prompt caching, an evaluable drafter, and the design decisions I made along the way.
PrestaShop is the workhorse of EU e-commerce, and bolt-on AI is the obvious next move for catalog ops. Here is a minimal 5-file PrestaShop module pattern that wires Claude SDK into the back-office without touching core — production-ready, with the failure modes I have hit.
Most "AI feature" projects skip the eval harness — and then silently regress for weeks before someone notices. Here is the production pattern I use: fixture-based regression tests with tolerance bands, run on every prompt change and every Anthropic model update.
Prompt caching with Claude SDK is the cheapest 10× cost win you can deploy in a single afternoon — when you use it correctly. Here are the 4 patterns I ship in production, the cost math behind each, and the 4 cases where caching does not help.