The LLM Is a Function, Not a Friend
Before you write a single line of LLM code, install the one mental model that makes every other decision easier: it is a function, not a thing you are talking to.
Apr 13, 202612 min read3
Command Palette
Search for a command to run...
Series
AI for Builders
Four Ways to Call a Model, and Which One You Actually Want
curl, SDK, streaming, async. Four ways to call an LLM, one shape that beats the rest in production, and the ones everyone reaches for at the wrong moment.
Apr 13, 202611 min read2Stop Parsing Strings: Structured Output From LLMs
Asking nicely for JSON is not a plan. Here is how to get structured output from an LLM that your downstream code can trust, with schemas, real errors, and the moves that scale.
Apr 13, 202611 min read1Temperature, Top-p, Seeds: the Knobs You Actually Touch
Most developers pick defaults for temperature and top-p and never revisit them. Most defaults are wrong for most calls. Here is what each knob actually does and when to touch it.
Apr 13, 202610 min read1System, User, Tool: the Real Prompt Hierarchy
Every LLM call has three kinds of messages: system, user, and tool. They disagree all the time. Here is who wins, why, and how to design prompts that don't get overridden by the next user turn.
Apr 13, 202610 min readPrompts Are Code, Not Config
Most teams treat prompts like magic strings. Then they wake up to a regression nobody can explain. Here is how to treat prompts like code — versioned, templated, tested — without making a career of it.
Apr 13, 202610 min read
