Conclusion First

Outcome: the user has said, in their own words, what the paper would conclude if everything worked, what its single idea is, which part is most likely to fail, and what they will do now. The agent asks and reflects; it does not score.

Scope: long-horizon work only, per the repository's AGENTS.md. A near-term deliverable does not get this gate.

Adapted from the essay's conclusion-first test and its sections on killing projects and starting with the part most likely to fail.

The conversation

Ask these in order, one at a time, and stop early when the answer settles the matter.

If every experiment turned out exactly as you hope, what does the conclusion say? Wait. If the answer is a number going up, ask: what changes about how people think or build? If nothing, say plainly that this may be sound science but not the important project, and ask whether to reframe or stop.
What is the one idea, in one sentence? If the sentence needs an "and", say that this is two projects and ask which one this is.
Which part is most likely to fail? What is the smallest thing you could try in days that would tell you? If the user has spent weeks on the part they already knew how to do, name that as the first finding.
Only when the project is already running: is the idea failing, is it working but not mattering, or has something more important appeared? For the third, add that new ideas always look better than the one you have lived with, and ask what specifically makes the new one more important.
What will you do? Continue, kill, pivot, or de-risk first. If killing, what is salvaged: a workshop note, a blog post, or an entry in the taste journal.

When a project is ending well, add one more question: is there an obvious experiment, domain, or objection a reader will wish you had addressed? If yes, it belongs in this paper. Small follow-ups can stay open; they are how others join.