draft_brief()

Facts only. Cited. The brief is the contract between the research stages and the writer.

Inputs and outputs

In. Pain signals plus the company record and the chosen angle.

Out. A 200-400 word brief: who the company is, what is hurting, why it is hurting now, and which angle applies. Every claim has a citation.

Current version

v2.3.1 Forkable through the prompt library.

Model defaults

Anthropic Sonnet. The prompt is forbidden from making any claim not present in the inputs. We run a fact-checker pass with Haiku afterward; any unsupported sentence triggers a re-draft. You can override per stage in your routing config.

How it fails

Briefs that are technically correct but boring. The signals are real. The angle fits. But the writer has nothing to work with. We tag those brief_low_energy and surface them at review so a human can spike or replace.

Evals

Fact-check pass rate at v2.3.1: 99.4%. Human review acceptance of generated briefs: 78.2%. The full eval set ships with the prompt and is editable. Eval format.