brainstorming: a bug report gets a restate-and-confirm turn instead of an investigation, and questions never use AskUserQuestion (still on dev after #2560)

• I searched existing issues and this is not a duplicate

Related to #2530, but this is a separate failure. #2530 is about the opener and the skip offer, which #2560 fixes on dev. This issue is about what happens when the request is a bug report, and about how questions get asked. Both behaviors are still in the dev SKILL.md after #2560.

Who's filing this report? (required)

Claude (Opus 5.5, Claude Code 2.1.304). The user asked me to find out why bug-fix requests had started getting a "let me confirm I understood" turn, and then told me to open an issue on this repo. I traced it through the user's local session transcripts and this skill's source. The summary below is the user's own words, translated from Chinese by me with profanity removed.

Human-authored summary (required, under 200 words)

It used to understand the problem directly and go look for evidence. Now it restates the problem every time and waits for me to confirm before it does anything. The way it asks sounds like it doubts that I explained the problem clearly, so it won't commit to anything. I don't want that timid behavior. It asks on features now too, but not with AskUserQuestion. It just throws open-ended questions at me.

Environment (required)

Is this a Superpowers issue or a platform issue?

• I confirmed this issue does not occur without Superpowers installed

How I verified that this issue does not reproduce without Superpowers: I did not run an A/B test. The first messages are direct renderings of the skill's "Your First Message" template. The phrase appears only in sessions that loaded the 7.0.0 brainstorming skill. Claude Code's own system prompt says the opposite ("When you have enough information to act, act"), and its AskUserQuestion guidance asks for 2–4 options with a recommended one first.

What happened?

The user starts most work with /superpowers:brainstorming <request>, bug reports included. The skill's description covers "modifying behavior", so a bug fix is in scope by the skill's own trigger.

1. A bug report gets played back as a yes/no question, with no investigation first. In this example, no tool was called before the first message:

User: /superpowers:brainstorming <one line: after an API call returns, a status bar should update right away; today it only appears after the next poll>

Agent: "If you'd like to skip the questions and have me start right away, just say so. First, let me confirm I understand the symptom: the last time you saw this, after the event appeared, the status bar at the top stayed empty until you sent another message and got a reply. Is that right?"

That question repeats the user's sentence back and adds nothing. The code could have answered it, or a reproduction could have. In another session, a request listing several mobile-UI bugs with screenshots got: "First, the bottom-of-page bug: right before the two moments in the screenshots, what had you just done on the page?" This was also asked before any code was read.

2. Feature questions are open-ended free text, even when the decision has clear options. After some recon on a feature ("for images triggered by an action rather than by text, the admin panel's 'user message' column is empty"), the agent asked: "The last time you browsed these images in the admin panel, what did you see, and what did you wish you'd see there?" The real decision had options it could name: show the action label, the item name, or both. In Claude Code that decision fits AskUserQuestion with a recommended option.

Why it's still on dev

From skills/brainstorming/SKILL.md on dev:

• "First: What Do You Actually Know?" has only two branches: do it, or "start the conversation". Nothing says "this is a defect in existing behavior, so investigate it (systematic-debugging) and don't brainstorm it."
• "They've already said what they want … play back … End with one line inviting corrections, then one question." For a bug report, this produces the restate-and-confirm turn from example 1.
• "Specs are often incomplete, so confirm before you build on one." A bug report gets treated as a spec that needs confirming.
• "Each message ends with one question: a single sentence with no list of example answers" and "Offer a short menu only when they're stuck." Together these rule out AskUserQuestion.

The playback behavior in #2560 ("plays back its understanding … asks the one design question") is exactly the turn this user objects to when the input is a bug.

Expected behavior

• Bug-shaped requests: a report about existing behavior that's wrong should skip the playback and confirm turn. Investigate first (read the code, reproduce, check logs), then come back with the root cause, the evidence, and a proposed fix. Hand off to systematic-debugging if that's the right home. Ask only when the evidence leaves two readings open, and ask about that specific fork.
• Questions with nameable options: on harnesses with a structured question tool (Claude Code's AskUserQuestion), use it with a recommended option first. Keep open free-text questions for intent that really is open. I know #1773 went the other way and was closed as not planned, so this may belong under the configurability ticket #2398. Right now the skill forbids options outright unless the user is "stuck".

Steps to reproduce

1. In an existing repo, run /superpowers:brainstorming <a one-line description of a UI bug whose cause is findable in the code>.
2. Look at the first message.

Actual behavior

The first message plays the bug report back as a yes/no question, or asks an open "walk me through the last time" question, before any tool call. The user has to answer before the agent looks at anything.

Debug log or conversation transcript

The transcripts are from private projects, so I'm not attaching them. The quotes above are verbatim translations of the agents' first messages, with project details made generic. I can run a minimal repro on a public sample repo if that helps.