brainstorming: v6.0.0 platform-neutral bootstrap reframe makes Claude Code use AskUserQuestion (regression vs 5.1.0)

[obra]

• I searched existing issues and this is not a duplicate

Environment (required)

Field	Value
Superpowers version	6.0.0 (also checked against the v5.1.0 tag)
Harness (Claude Code, Cursor, etc.)	Claude Code
Harness version	2.1.178
Your model + version	Claude Opus 4.8 (claude-opus-4-8[1m])
All plugins installed	superpowers (6.0.0, local working tree), episodic-memory, linear, context7, claude-in-chrome, superpowers-chrome, superpowers-lab, superpowers-developing-for-claude-code, plugin-dev, github-triage, primeradiant-ops
OS + shell	macOS (Darwin 25.2.0), zsh

Who filed this: Authored by Claude (Opus 4.8) running in Claude Code, in a working session with Jesse Vincent (@obra), the project's author/maintainer. Grounded in a real git-history investigation of this repo (not reasoned from documentation).

Is this a Superpowers issue or a platform issue?

Superpowers issue — the regression is in Superpowers' own bootstrap content (skills/using-superpowers/SKILL.md), so the "reproduce without Superpowers" check is N/A.

What happened?

On 6.0.0, the brainstorming skill on Claude Code now reaches for the native AskUserQuestion tool to ask clarifying questions / present option sets. On 5.1.0 it asked in plain conversational text, one question at a time, as the skill intends.

Root cause

The brainstorming SKILL.md is effectively unchanged between v5.1.0 and v6.0.0 (the only diff is visual-companion timing). The behavior change comes from the v6.0.0 "platform-neutral" reframe of the Platform Adaptation section in skills/using-superpowers/SKILL.md:

- Skills use Claude Code tool names. Non-CC platforms: see references/copilot-tools.md ...
+ Skills speak in actions ("dispatch a subagent", "create a todo", "read a file")
+ rather than naming any one runtime's tools. For per-platform tool equivalents ...
+ see [claude-code-tools.md], [codex-tools.md] ...

• 5.1.0: "skills use Claude Code tool names" anchored the model to read brainstorming's prose literally — "Present options conversationally" (SKILL.md:78), "Use the terminal for … A/B/C/D text options" (SKILL.md:154). No tool is named there, so it asked in text.
• 6.0.0: "skills speak in actions … map them to your native tools" reframes every skill verb as a mappable action. "Ask clarifying questions / present multiple-choice options" now resolves to Claude Code's native interaction tool, AskUserQuestion. The skill's own "Prefer multiple choice questions when possible" (SKILL.md:71) and "Multiple choice preferred" (SKILL.md:136) make AskUserQuestion the obvious mapping target — overriding the explicit "present conversationally / use the terminal" intent.

Two corroborating points:

1. No explicit prohibition ever existed. A pickaxe over full history shows the literal string AskUserQuestion appears nowhere in 5.1.0 except RELEASE-NOTES.md. The protection on ≤5.1.0 was structural (skills written in literal CC tool names + prose, bootstrap said "use them as written"), and the v6 reframe dissolved it.
2. The v6 authors mapped these exact phrases to AskUserQuestion for Kimi — deliberately. The new .kimi-plugin/plugin.json skillInstructions say: "When a Superpowers skill says to ask the user, ask clarifying questions, ask one question at a time, present multiple-choice options, use the terminal for a question … call … AskUserQuestion." On Claude Code the same routing now happens implicitly and unintentionally through the genericized bootstrap, with nothing countering it. docs/porting-to-a-new-harness.md (new in v6) never discusses the "ask the user" case, so this was an unconsidered side effect of the platform-neutral work.

Steps to reproduce

1. Install Superpowers 6.0.0 on Claude Code.
2. Start a fresh session and send a build/design request (e.g. "Let's make a React todo list") so brainstorming auto-triggers.
3. Observe the clarifying-question phase.

Expected behavior

Clarifying questions presented as plain conversational text, one at a time, in the terminal — matching the skill's "present conversationally" / "use the terminal for text options" intent (the 5.1.0 behavior).

Actual behavior

The model invokes the native AskUserQuestion structured-choice tool instead of asking conversationally.

Related prior issues (searched, not duplicates)

These all requested the opposite — adopting AskUserQuestion in brainstorming — and are closed: #388, #114, #745 (closed COMPLETED), #1197 (closed NOT_PLANNED). None describe the v6.0.0 regression where the tool gets used unintentionally as a side effect of the platform-neutral bootstrap. If the project's settled position is "brainstorming asks conversationally, not via AskUserQuestion," this issue tracks restoring that on 6.x.

Possible fix directions (for discussion, not yet implemented)

• Add an explicit Claude-Code-scoped guardrail in skills/using-superpowers/references/claude-code-tools.md: asking the user is not a tool action — present questions/options as plain conversational text, one at a time; do not use AskUserQuestion. (Keeps it CC-specific; harnesses like Kimi still opt in.)
• Or an explicit "present as text, not a structured-choice tool" line in skills/brainstorming/SKILL.md near the "multiple choice preferred" guidance.

Open question for the maintainer: scope the suppression to Claude Code only, or suppress AskUserQuestion during brainstorming on every harness except those that explicitly opt in (as Kimi does)?