[Feature Request] Add critique skill for nine-dimension self-review of AI responses

[ZhangShiZhao-Beta]

• I searched existing issues and this has not been proposed before

What problem does this solve?

When AI agents produce substantive responses — particularly architectural designs, technical analysis, or complex recommendations —
they often contain quality issues that aren't obvious during initial review. I've experienced this firsthand: after receiving a
complete architectural design from the AI, I asked it to confirm the design, and it found no obvious issues. But when I prompted a
structured self-examination using critical thinking standards, the AI uncovered real problems it had missed — vague abstractions,
overlooked edge cases, and logical gaps.

This is different from code review (which checks code against plans/specs) and verification-before-completion (which validates
outcomes). There's currently no skill that systematically evaluates the quality of the AI's own response across multiple
intellectual dimensions — clarity, precision, accuracy, completeness, logic, etc.

Proposed solution

A critique skill that triggers a nine-dimension self-review of the AI's most recent substantive response, based on the
well-established intellectual standards from critical thinking theory:

1. Clarity, 2. Precision, 3. Accuracy, 4. Importance, 5. Relevance, 6. Completeness, 7. Logic, 8. Breadth, 9. Depth

Key design decisions:

• Manual trigger only (/critique) — not automatic, to avoid slowing every interaction
• Output only dimensions with problems found — no noise from passing dimensions
• Severity-ordered (P0/P1/P2) with classification tags ([error], [reword], [missing], [excess])
• Max 5 problems — keeps feedback focused and actionable
• Language-matched — follows current conversation language
• Waits for user direction — doesn't auto-fix, lets user choose what to address

What alternatives did you consider?

1. Using verification-before-completion — This verifies implementation outcomes against requirements, not the quality of a
   conversational response. Different scope.
2. Using requesting-code-review — This dispatches a subagent to review code against a plan. It doesn't evaluate the
   intellectual quality of an AI response.
3. Simply asking "review your answer" — In practice, the AI gives superficial self-reviews without a structured framework. The
   nine-dimension approach produces significantly deeper and more actionable feedback.

Is this appropriate for core Superpowers?

Yes. This skill benefits anyone using AI coding assistants regardless of project type, language, or framework. It applies equally
to architectural discussions, debugging analysis, planning, and general conversation. It's platform-agnostic and doesn't depend on
any specific tool or domain.

Context

I've been using this as a personal skill for Claude Code with strong results. It's particularly valuable after brainstorming
sessions and architectural design discussions where the cost of undetected quality issues is high.

[nsalvacao]

I want to add one data point here because I ran into the same underlying problem and built a personal critique skill for it.

First: I think the problem statement in this issue is real and well-framed. I have seen the same failure mode in practice: asking the model to “review” or “double-check” a substantive
answer often produces shallow reassurance, while a structured reasoning critique surfaces actual weaknesses such as vague abstractions, missed edge cases, weak trade-off handling, and
logic gaps.

What I have locally

I built a personal critique skill for this exact class of problem. It is meant to critique one reasoning artifact at a time and produce evidence-backed findings without auto-fixing.

It can target:

• the assistant’s prior answer
• a user-provided argument
• an ADR
• a plan
• a spec
• a pasted reasoning artifact

The core behavior is:

• only report findings with observable evidence
• severity-order the findings
• do not rewrite or auto-correct the artifact
• stop after the critique and wait for user direction

How my version differs from the proposal here

My current local version is not a drop-in implementation of this issue.

Main differences:

• this issue proposes a manual-only /critique; my current version is broader and can trigger implicitly
• this issue is focused on self-review of the AI’s most recent substantive response; my version critiques any single reasoning artifact in scope
• this issue proposes a 9-dimension review; my version uses the Paul/Elder-style standards plus an always-on Completeness heuristic
• this issue proposes max 5 problems; my version currently prioritizes and can report more than that
• this issue proposes language-matched output; my current version is English-only

So the overlap is strong at the problem level, but my implementation made slightly different product decisions.

Methodological reference

The reasoning-review method I used was based on the Paul/Elder intellectual standards tradition, with this as the external reference point:

• Foundation for Critical Thinking, “Universal Intellectual Standards”:
  https://www.criticalthinking.org/pages/universal.intellectual-standard/527

Why I’m commenting instead of opening a PR immediately

I do have something close to what is being proposed here, but I would not upstream my local version unchanged.

If there is maintainer interest, I’d be happy to adapt it to the narrower shape proposed in this issue:

• manual-first invocation
• the exact dimension set expected for core
• max 5 findings
• output format aligned with Superpowers conventions

I mainly wanted to confirm that I have the same need, that I built a working local version for it, and that there is some concrete prior art here rather than just a theoretical
suggestion.