extract_signals()

The stage where the product earns its name. Find the specific, dated, cited evidence that this company hurts in a way you can fix.

Inputs and outputs

In. Outputs of 01-04 plus the pack's signal taxonomy (typically 28-47 signals).

Out. Zero or more pain signals. Each signal has a type, a date, a confidence, a source URL, and a verbatim excerpt.

Current version

v0.9.4 → v1.0.0 ships next week. Read the diff. Forkable through the prompt library.

Model defaults

Anthropic Sonnet for the extraction call. Anthropic Haiku for the citation checker that verifies every signal is actually present in the source text. You can override per stage in your routing config.

How it fails

Hallucinated dates. Older versions of this prompt would generate plausible signals with invented timestamps. v0.9 introduced a hard citation gate: a signal without a verifiable excerpt is dropped, not softened. Drop rate at gate: 11.4%.

Evals

On a hand-labelled set of 1,200 signals across six packs: 91.8% precision, 76.2% recall at v0.9.4. v1.0.0 holds precision at 92.1% and lifts recall to 81.0%. The full eval set ships with the prompt and is editable. Eval format.