feat(leads): generate the pitch fields from a plain description of the goal

[ralyodio]

The campaign form asks for three precisely-scoped inputs, and the grounding guard refuses drafts that stray outside them. That's a lot to ask of someone who just wants to describe what they're doing — and the failure is silent.

A live campaign had the same sentence pasted into both the intro and the ask, a facts list omitting the URL the ask depended on, and five rejected drafts a tick for hours as a result.

Describe the goal, get the fields

Verified on that exact campaign's text:

In: "Anthony owner of threatcrush.com we want to generate leads for people to download our free whitepaper at https://threatcrush.com/get-whitepaper"

Out:

INTRO  I'm Anthony, owner of threatcrush.com, and I'm reaching out because
       we've created a resource that might be useful to you.
ASK    Download our free whitepaper at https://threatcrush.com/get-whitepaper
FACTS  - threatcrush.com offers lifetime access to real-time threat intelligence
       - We have a free whitepaper available at https://threatcrush.com/get-whitepaper

Then, against the guard: a draft using the URL passes [], and "We have 4200 customers" is still rejected.

The first fact came from reading threatcrush.com's own site description — the one extra fact available without guessing, and exactly what someone omits describing a goal in a hurry.

It restructures; it does not author

This is why it can be trusted with the job. Drafts are checked against its output, so a fact invented here isn't caught later — it's laundered into something a stranger is told is true. The prompt refuses new numbers, links, dates and credentials, and refuses to embellish ("30 years as a software engineer" must not become "three decades of industry leadership").

Any URL the person wrote is forced into the facts in code afterwards, not trusted to the model — dropping one recreates exactly the failure this exists to prevent.

Also

The two guard messages disagreed — one said "not in the campaign's declared facts", the other "the campaign never mentions". After #148 widened the grounding set, the first was no longer true.

Checks

• tsc --noEmit clean · 938/938 tests pass · build compiles
• generator exercised live against the real campaign goal

🤖 Generated with Claude Code

[github-actions]

vu1nz Security Review

0 finding(s) in PR #?

No security issues found.