Writing
The agent writes the diff. Plain code decides if it ships.
I let a coding agent take GitHub issues to merged PRs. The trick was not a better prompt. It was keeping the agent away from every decision that reaches production.
Read →Your checkout is wrong in Kuwait
A lot of checkout code assumes every currency has two decimals. The Kuwaiti dinar, the Bahraini dinar and the Omani rial have three, so "multiply by 100" charges a tenth of the price and nothing crashes.
Read →I gave my AI a workflow instead of a chat window
An AI in a chat window is a brilliant contractor with amnesia. I stopped being its memory and moved it into my repo, where state lives in files and it runs one phase of my day at a time.
Read →How to handle a production incident without making it worse
When something breaks in production the instinct is to dive into the code. The right first moves are calmer than that. Stabilize, communicate, then find the cause.
Read →Stop parsing LLM text. Use tool calls.
Asking a model for JSON and parsing the text breaks in production. Define a typed tool, force the model to call it, and get validated structured data back.
Read →Cut your LLM bill with prompt caching
Most AI features resend the same large prompt every call and pay full price for it. Cache the stable prefix and pay a fraction on every reuse.
Read →