AI Engineering Insights
Agent orchestration, context management, domain intelligence — practical notes you can use today.
Stop chatting, write one prompt — Using LLMs Well (5)
Firing one line at a time — "fix this," "oh, and that too" — feels natural, the way you text on WhatsApp. It is also quietly expensive, because an LLM has no memory and re-computes the entire conversation on every turn. Here is why fragmented commands cost double, how far prompt caching saves you and where it breaks, and how to say it all in one complete prompt.
Latest
Don't Trust "Done" — Using LLMs Well (3)
The agent says "all fixed, tests pass." Then it breaks in production. This isn't the agent lying — it's a structural consequence of autoregressive generation. Here is a technical analysis of why you can't trust a completion report, and a step-by-step way to ground claims in evidence.
A Shorter CLAUDE.md Gets Followed Better — Using LLMs Well (2)
It feels like more rules means better adherence. The opposite is true. Here is why an always-loaded instruction file gets ignored the longer it gets — fixed cost, attention dilution, lost-in-the-middle — how to prune it like code, and what RoutineCode does instead of "always load."
Precise Requests Are Token Economics, Not Politeness — Using LLMs Well (1)
Ask an agent to "look at the login stuff" and it wanders the whole codebase. That is not laziness — it is how LLMs work. Starting from two facts (attention is zero-sum, context is a fixed budget), here is why precise requests are a performance optimization, and how RoutineCode implements the same principle.
Get new posts first
Dev-workflow and RoutineCode updates by email. No spam.