/debate <idea> → two agents with opposing briefs in parallel → verdict on evidence quality
A single researcher investigating an idea mostly finds evidence for their first impression — motivated search is invisible from the inside. Splitting the search into two opposing briefs makes the bias work for you: each side hunts hard, and only evidence that survives the other side's case earns a place in the verdict.
The input is a crisp idea — problem, person, wedge. If it's still fuzzy, clarify it first (that's /talk-about-an-idea's job); debating a vague idea just produces two piles of unrelated links.
The two briefs
Launch both agents in a single message so they run concurrently. Neither gets a persona or a prior — just an opposing brief. Each brief includes the idea verbatim, the research discipline quoted verbatim (an agent can't follow rules it never sees) — every claim links to its source, evidence stays separate from inference, findings carry dates — and /research's method as the toolbox: go wherever the answer leaves a first-hand trace, and treat secondary accounts as leads, not evidence.
- For — the strongest honest case that this is worth building: demand behavior, people paying for inferior substitutes, recurring complaints about the status quo, gaps competitors leave open.
- Against — the strongest honest case that it isn't: dead or struggling competitors who tried, users content with the workaround, structural moats, evidence the pain is mild or shrinking.
"Strongest honest case" means advocate, don't fabricate — evidence each side searched for and couldn't find is part of its case, and each returns that too.
The verdict
Both sides will always find something, so never count evidence — weigh it. Observed behavior (paying, switching, building workarounds) outranks complaints; complaints outrank opinion pieces; recent outranks stale; independent sources corroborating outrank one loud voice. The verdict rests on which case's best evidence survives the other case.
## Debate: <idea>
**Verdict:** <worth building / not worth building / hinges on <the one unresolved question>>
## The case for
<its surviving evidence, sourced>
## The case against
<its surviving evidence, sourced>
## What would change the verdict
<the concrete evidence that would flip it, and where to look>
Close with the verdict and stop — whether to proceed is the user's call. A "worth building" flows into /grill-me → /to-spec; a "hinges on" names the experiment to run first.