qa-reviewer-sparring
English | 繁體中文
A sparring partner for quality-assurance reviewer judgment.
Drop in an evaluation scenario — a chunk of a self-assessment report, a site-visit slide, or an item you're unsure how to question — and the skill takes the role of a seasoned reviewer. It doesn't grade you, correct you, or flatter you. It catches the question you asked, then pushes on the side you didn't ask about: "How do you know it wasn't something else?", "Where did the ones who didn't make it go?", "Is this evidence enough to draw that conclusion?" After a few rounds, it reflects your questioning patterns back to you and names your systematic blind spot.
It is a universal Agent Skill — one SKILL.md file that runs on Claude Code, Codex, Gemini CLI, opencode, and the Claude desktop app, plus any chat interface via paste. It is not an agent and needs no backend.
Disclaimer. This skill is a training aid for practicing reviewer judgment on synthetic cases. It is not an evaluation, audit, or accreditation instrument, and it is not intended to review or assess any real institution's report, documents, or submission. Its output has no official standing and should not inform any real review decision. The skill is provided "as is," without warranty of any kind; any use of it, and any consequence of that use, is the sole responsibility of the user. See also the LICENSE.
How it works
ONE SPARRING SESSION
─────────────────────
you trigger it
│
▼
it states what it does / doesn't do (no scoring, no certifying)
│
▼
"what did you bring?" ──┬── you paste real or sample material
├── nothing? it hands you a synthetic case
└── a progress snapshot? it resumes from there
│
▼
YOU answer first ── you write your probe before it shows its hand
│
▼
┌─────────────────────────────────────────────────┐
│ catch it names the angle you just used │ each round,
│ ↓ (no rating) │ one lens
│ push back it asks about the side you │ at a time
│ DIDN'T ask about │
└─────────────────────────────────────────────────┘
│
▼
blind-spot read + a progress snapshot (plain text you keep)
ACROSS SESSIONS → COMPLETION
───────────────────────────────
session 1 ──▶ snapshot ──▶ paste back ──▶ session 2 ──▶ snapshot ──▶ ...
│
once all six skill strands are reached ──────────▼
completion threshold
│
▼
a questioning-style profile report (a profile, not a score)
What you're training
This is a desk-review trainer. Every case is a piece of written material — a self-assessment passage, a SWOT block, a promotion-weighting table — and the skill drills the judgment you bring to reading a report: reaching for the evidence the page doesn't show, not reacting to a person in the room. (An on-site interview mode — live, in-person questioning where you can't go back and re-check the evidence — is a separate skill, not built yet.)
Six skill strands
What separates a reviewer from a novice, and from a generic AI, isn't knowing the standards — anyone can look those up. It's asking about the side that didn't happen. The skill trains six strands toward that:
| # | Strand | The question it forces |
|---|
| 1 | Causal Exclusion (Lens 1) | Could the result be something else, not this initiative? |
| 2 | Assume-and-Falsify (Lens 2) | Assume the claim is overstated — what evidence would overturn that? |
| 3 | Failure-Side Check (Lens 3) | Where did the ones who didn't make it go? |
| 4 | Cross-Source Corroboration | Do the report, the data, and the narrative actually line up? |
| 5 | Evidence Sufficiency | Is this evidence enough to draw that conclusion? |
| 6 | Bird's-Eye Orientation | What's the whole picture before I dig into one number? |
The first three are counterfactual lenses — they puncture a polished report. Both humans and AI reflexively reach for the familiar Lens 1 and silently skip Lens 2; that skipped reflex is the single most common gap this tool exists to correct.
Read the situation first, then pick your stance
The cases aren't a gotcha question bank. They deliberately span five real review situations, because a reviewer who can only nitpick fails just as badly as one who only goes easy. The skill makes you judge which situation a report is in before deciding how hard to push:
