Module 10 of 16
Model interfaces
The contract between your code and a model: tokens and context as a budget, structured output and tool calls as a protocol, and prompts as versioned, tested artifacts.
Everything a model does for you passes through a narrow interface: a sequence of tokens in, a sequence of tokens out. This module treats that interface the way you would treat any other API contract, with its units, its failure modes, and its versioning.
The first lesson works out what a token is, what a context window costs in memory, time, and money, and why prompt order decides how much of it is cached. The second covers getting machine-readable output: how constrained decoding masks tokens so the output always matches a schema, what it cannot guarantee, and how a tool call travels from the model’s proposal through your validation to an effect. The third treats prompts as code: content-addressed versions, regression suites with paired statistics, and release gates. Afterwards you should be able to budget a context window, design a tool interface the model cannot misuse into a bad effect, and block a prompt change that quietly breaks one slice of traffic.
Lessons
- 10.aTokens, context, and costWhat does a token cost in compute, memory, and money, and how does prompt layout change the bill?18 min
- 10.bStructured output and tool callsHow do you get output a program can rely on, and how should a tool call travel from the model to an effect?19 min
- 10.cPrompts as versioned codeHow do you change a prompt or a model without silently breaking part of your traffic?12 min