Prompts

What a prompt is

A prompt is a function whose body is a string template, whose return type is a typed value (or an Option<T> / Result<T,E> if the parse might fail), and whose effect row carries the LLM call’s cost, latency, and confidence.

prompt summarize(text: String) -> String uses llm_call:
    "Summarize the following in one sentence: " + text

How interpolation works

Variables in the prompt body interpolate. Strings concatenate with +. There is no implicit string conversion — non-string values must be converted explicitly:

prompt classify_priority(score: Int) -> String uses llm_call:
    "Score is " + score.to_string() + ". What priority?"

Typed return values

struct Decision:
    refund: Bool
    reason: String

prompt decide(ticket: String) -> Decision uses llm_call:
    "Given ticket: " + ticket + "\n\nReply as JSON {refund, reason}."

The runtime asks the model to emit JSON matching the Decision schema and parses the response into a Decision value. Parse failure is a typed error.

For per-struct decoders the compiler emits at codegen time, see the slice that landed this surface (Phase 20n-C).

Multi-message prompts

prompt ask(question: String) -> String uses llm_call:
    system: "You are a careful, terse assistant."
    user: question

Each role gets its own template. The runtime renders the messages, sends them, parses the response.

Provider routing

Which LLM serves the prompt is decided by the typed model-routing substrate at deploy time, not by hardcoding the provider in the prompt body. See docs/internals/effect-spec/13-model-substrate-shipped.md.

Budgets

A prompt’s cost dimension flows into the enclosing agent’s budget. A prompt that runs ten times in a loop counts ten times.

Replay

Every prompt invocation is recorded with input, output, model, latency, and cost. corvid replay reproduces the run from the recording without hitting the provider. corvid eval --swap-model re-runs the prompt against a different model and diffs the result against the baseline.