adversarial-review - SKILL.md Agent Skill

name: adversarial-review description: Run an adversarial peer-review swarm on a paper draft, report, or manuscript by spawning parallel persona subagents (hostile theorist, experimentalist, statistician, journal editor, citation auditor, figure critic) and synthesizing their critiques into a ranked fix list with suggested defense experiments. Use when the user asks for a paper review, referee simulation, desk-reject check, pre-submission audit, citation audit, figure audit, novelty attack, or wants feedback on a draft report before submission or internal circulation.

Adversarial Review

Use this skill to stress-test a draft before it reaches real reviewers. It spawns a small swarm of specialized persona subagents in parallel, each attacking the manuscript from one angle, then synthesizes the critiques into a prioritized fix list.

This is the opposite of a friendly read-through. The goal is to surface desk-reject-level problems, unsupported claims, weak statistics, and citation hallucinations before submission — not to reassure the author.

When to use

Pre-submission audit for a journal (PRX, JCAP, JOSS, etc.) or conference.
Internal circulation of a research report or technical memo.
Response drafting for a revise-and-resubmit where the author wants to predict reviewer pushback.
Any draft the user wants "torn apart" or "reviewed hard."

If the user wants a friendly copy-edit pass, this is the wrong skill — suggest a plain editing request instead.

Inputs to confirm

Ask only for what is missing:

path to the draft (default: report.md in the current directory)
journal or venue the draft targets (affects the Journal Editor persona)
any sections the user wants emphasized or skipped
whether figures live alongside the draft and where (plots/, figures/, etc.)
whether a bibliography file or a references section should be audited
whether Korean translation of the final summary is expected

If nothing is specified, default to:

draft: report.md in CWD
venue: "general physics/ML journal" (Journal Editor persona stays generic)
figures: plots/ if present, else skip Figure Critic
Korean translation: no, unless the user previously asked

Core workflow

1. Locate and stage inputs

Before spawning any agents:

Read the draft once end-to-end to understand the claim structure.
Enumerate figure paths and bibliography entries.
Produce a compact context packet that every persona will receive: draft path, venue, figure list, bibliography location, and any sections to emphasize or skip.

Do not let each persona re-read and re-summarize the whole draft from scratch — give them the packet plus the draft path.

2. Spawn personas in parallel

Spawn the persona subagents using the Agent tool, in a single message so they run concurrently. Use references/personas.md for the exact persona briefs; each persona gets:

the context packet
its own persona brief (role, what to attack, what to ignore)
an output contract (see below)
an explicit length cap so critiques do not bloat

Default persona set (skip any the draft does not support):

Hostile Theorist — attacks novelty; uses reference-search to surface prior art the draft may have missed.
Experimentalist — attacks reproducibility, missing ablations, undisclosed controls, hyperparameter sensitivity.
Statistician — attacks error bars, significance claims, baseline fairness, multiple-comparison exposure.
Journal Editor — predicts desk-reject reasons for the target venue; checks scope fit, framing, required sections.
Citation Auditor — verifies every non-trivial citation resolves to a real paper via reference-search; flags hallucinations and mischaracterizations.
Figure Critic — checks figure paths resolve, captions match the data, axes are consistent, cutoffs are applied uniformly, and no figure is referenced but missing.

3. Output contract (per persona)

Every persona must return a markdown block with this exact structure so the meta-editor can merge them deterministically:

# <Persona Name>

## Severity-classified findings
- **DESK-REJECT**: <finding> — <evidence pointer: section/figure/citation>
- **MAJOR**: <finding> — <evidence pointer>
- **MINOR**: <finding> — <evidence pointer>

## Suggested defense / fix
- <action the author should take>

## Confidence
<high | medium | low> — <one sentence why>

Personas that find nothing at a severity level simply omit that bullet. Never fabricate findings to fill the template.

4. Meta-editor synthesis

After all personas return, run the meta-editor pass described in references/meta-editor.md. It:

deduplicates overlapping critiques
ranks by severity (DESK-REJECT > MAJOR > MINOR)
attaches a proposed defense experiment or textual fix to every DESK-REJECT and MAJOR
drops low-confidence findings that no other persona corroborates

Meta-editor output is the canonical summary.md the user reads first.

5. Write outputs

Write all artifacts under outputs/review/<YYYY-MM-DD-HHMM>/ relative to the draft's directory. Keep individual persona files alongside the synthesis so the user can trace any claim in summary.md back to its source.

outputs/review/2026-04-18-1430/
  summary.md                 # meta-editor synthesis — read this first
  hostile_theorist.md
  experimentalist.md
  statistician.md
  journal_editor.md
  citation_auditor.md
  figure_critic.md
  context_packet.md          # the packet given to every persona (for reproducibility)

6. Report and stop

Show the user the path to summary.md, the count of findings by severity, and the three highest-severity items inline. Do not start fixing the draft. The user decides which findings to act on.

Integration

See references/integration.md for how this skill wires into the rest of the collection:

reference-search is the citation and prior-art backbone for Hostile Theorist and Citation Auditor.
research-report produces the report.md and figure layout this skill consumes; reports that follow its conventions are easiest to review.
Korean translation of summary.md follows the user's standing convention when requested: a Sonnet subagent produces summary_ko.md next to the English file.

Guardrails

Never modify the draft. This skill is read-only with respect to the manuscript.
Never silently skip a persona. If a persona cannot run (e.g. no figures to critique), say so in summary.md under "skipped".
Never accept a persona output that invents findings not grounded in the draft, figures, or bibliography. If a persona returns unsupported claims, rerun that persona once with an explicit "ground every finding in a section, figure, or citation pointer" instruction; if it fails again, mark its section as unreliable in summary.md.
Never fabricate a citation when checking citations. If reference-search returns nothing for a cited title, flag it as "not found via OpenAlex — manual verification required" rather than guessing.
Never push the review outputs. Writing under outputs/review/... is fine; git operations are out of scope.

Resources

references/personas.md — the six persona briefs with role, attack surface, ignored concerns, and output contract.
references/meta-editor.md — dedup, severity ranking, and defense-experiment authoring rules.
references/integration.md — wiring notes for reference-search, research-report, and Korean translation.