#2950 stagecraft skill: getting started asks every question at once; it should ask one at a time

The getting-started walkthrough asks all its questions at once and opens with a wall of text. It should ask one question at a time, the way swamp-getting-started does.

What happened

Seth ran the walkthrough in a fresh demo repo on 2026-10-02, from stagecraft on origin/main at 9b7f2577e. He expected "the getting started flow to ask one question at a time and instead it still gave me a lot to read at once."

Cause

The skill tells the agent to batch:

• references/getting-started.md:69 (State 1, goal_understood): "Ask everything in one message, and say what you will assume for each", followed by five numbered questions.
• references/authoring.md:54 (State 1, interviewed): "ask everything in one message, not a drip", followed by five numbered questions.

The batching came from #2818's issue body ("Batch questions: one round, not a drip"). That body applied a rule meant for agent work sessions reporting to a maintainer to a product's conversation with its users. #2931 then inherited it. In addition, getting-started's "Before starting" shows a 7-item checklist, and the walkthrough opens by explaining concepts. The first reply is long before the person has said anything.

Fix

• getting-started.md
  • One question per turn, in order. Each question carries its default, so a short answer works: "Who takes part? I'll assume an agent drafts and a person approves before it goes out."
  • Reflect each answer back in a line before asking the next.
  • Stop asking once there's enough to pick an example. Use defaults for the rest, and say so.
  • Make the opening two or three sentences plus the first question. Replace the 7-item checklist with one line naming the steps ("we'll go step by step: your process, then the factory, a look at it, and a test run"), or show progress as each step completes. Move the concept explanation to the point where it's needed, briefly.
  • Keep each later state's messages short too: one action or question per turn, with a Verify that the person can see.
• authoring.md's interview (the path for someone making another factory): the same one-question-at-a-time rule. It can move faster, since the person knows the terms, but it never shows a numbered questionnaire.
• SKILL.md: make sure nothing tells the agent to batch questions to the person anywhere in authoring or getting started.
• Eval: add a case to the eval suite (#2912 is in progress; coordinate with it). Assert that the first reply to "set up a factory for our post-incident reviews" asks exactly one question and stays short: a regex count on question marks, plus a length bound, plus an llm rubric. Also assert that a two-turn exchange asks the next single question. If #2912 hasn't landed when this merges, give #2912 the case to add.

Done when

• In a fresh repo, the first reply to a getting-started prompt is a short greeting and one question. Each later turn asks one question or takes one step.
• The person is never shown a numbered list of questions to answer.
• Record the acceptance transcript in the PR: the first three turns.
• The eval case exists, here or handed to #2912.

Out of scope: changing the walkthrough's states or what they produce. This is only about how it talks.