06 WRITING
Notes from production: putting language models behind real features without losing the reliability habits a decade of platform work beat into me. I started writing these up in June 2026; the practices come from years of shipping.
- I broke my own prompt-caching rule and paid 24 seconds a request
- Getting reliable JSON out of an LLM in PHP
- How to tell if a prompt change helped
- Cost per request, logged next to latency
- Prompt injection in RAG starts with the documents you indexed
- Calling an LLM API from PHP without taking the page down
- A 200 you can't use is the new LLM failure mode
- Evals before vibes
- RAG over a decade of internal docs
- Prompt injection is just untrusted input
- Structured outputs deleted half my parsing code
- Where the money goes at campus scale