# Prompt for one search front

Five fronts run in parallel, each in its own fresh context, one domain each, none seeing the
others. One front works by analogy: a sector where the problem is already solved another way
and the new kind of product is entering. Fill the brackets and send each front this prompt.

```markdown
**Starting state:** <the project in three lines: what the product does today, what has been
measured, the vision in one line>. Filters approved by the founder on <date>:
1. <filter 1, with the worked real case>
2. ...
You do not see the other fronts.

**Target state:** a report for the <domain> front (or: the analogy front, a sector where this
problem is already solved another way and this kind of product is entering), with:
- candidates that pass every filter, one card each, in the format below;
- rejected candidates, with the filter each failed and the reason in one line;
- notes: findings that didn't become candidates.
Card format: <paste wedge-card.md>.

**Scope:** web search to find URLs, then open every page you cite. At most <N> searches on
this front. Write no files; return the whole report in your reply.

**Don't:**
- Use what the customer already pays, or the rule that obliges them, as a filter. That is
  evidence.
- Count a buyer from a regulatory authorisation.
- Sum or cite a number seen only in a search snippet. Label it `secondary`.
- Invent a customer, an incident or a quote.

**Done when:** every candidate has evidence with an opened URL, and every rejection names its
filter.
```

Why each rule:

- **Blind fronts** keep one front's framing from steering the others, so duplicates found
  independently mean something when you merge.
- **A search cap** keeps one front from using the whole budget and leaving the map lopsided.
- **Return the whole report.** Save it as is; the founder will ask for the detail behind a
  card later, and the card's Origin field points at it.