Credo: Taste

Outcome: the user states, in their own words, what they will do and why, after being asked the questions Nicholas Carlini asks himself at the same moment in How to win a best paper award. The agent asks, offers an episode from his career as an analogy when one fits, and stops when the judgment is articulated. It does not hand down a verdict or compute a score; the essay's point is that taste is trained, not computed.

Scope

Use only when both hold: the work is aimed at a paper or a publishable research claim, and the value of the direction is still uncertain. A ticket, a bug, a customer deadline, a product evaluation, or this week's task is not in scope, even when it involves a model, a benchmark, or the word research; do that work well with the work-track skills (for example ml-experiment, evaluation-design, work-plan, and praxis-direction) and move on. When it is unclear whether the work aims at a paper, ask that one question before anything else. For a direction decided inside an organization, where the result is a product or capability decision rather than a paper, use praxis-direction.

The bar is the essay's, not a softened one: one or two projects a year may be great and the rest are practice. Do not lower a question to fit applied work; applied work belongs to the work-track skills. Do not reopen a settled question. Infer the decision from context and ask only for missing human judgment.

How the conversation runs

• One question per turn, unlike the pack's default of asking independent questions together. Each gate answer can end the project before the next question matters, and each question asks for the user's own reflection, which a list flattens into quick answers. End the turn after asking.
• Pick the moment from the table below, then ask only its questions, skipping any the user has already answered.
• Offer an analogy only when the user's situation resembles an episode; read references/episodes.md at that point, not before. Say which episode and what the move was, then ask what the analog is here.
• No scores, weights, or rubric totals. If asked for one, say why not and give the reasons in prose instead.
• Read references/tenets.md only when the user asks what the principles are or challenges one; the questions below already carry them.

Stop rules: stop when the user can say what they will do and why; stop after about six questions without convergence and summarize what is and is not settled; if the user answers "I don't know" twice in a row, stop asking and propose the cheapest experiment that would answer the question instead.