Seven-Pass Adversarial Review

Source. This skill is adapted from pedrohcgs/claude-code-my-workflow. The seven-lens taxonomy (Abstract / Intro / Methods / Results / Robustness / Prose / Citations), the parallel-subagent-then-synthesizer topology, and the 80 / 90 / 95 quality-threshold scoring are Pedro H.C. Sant'Anna's; this fork tunes the journal targets to marketing (MKSCI / JMR / JCR / MS) alongside the econ / finance ones.

Spawns seven independent reviewers, each focused on a single lens, then synthesizes their findings into one prioritized revision plan.

Why seven passes? A single-agent review blends lenses and softens each one. Seven forked subagents each approach the paper with full context for their own lens, then a synthesizer resolves conflicts, deduplicates issues, and scores the manuscript against quality thresholds (80 / 90 / 95).

When to pick this over `/review-paper` or `/review-paper-light`: This skill costs roughly 7x more tokens than /review-paper-light and ~3x more than /review-paper. Use it when the paper is submission-ready, R&R-stage, or about to go to a top marketing or econ journal. For early drafts and iterative work, the lighter reviews are the right tool.

Inputs

• $0 — manuscript path (.tex, .qmd, .Rmd, .md, or .pdf). Required. Manuscripts typically live under <OVERLEAF_ROOT>/<PROJECT_SUBDIR>/.
• --journal — optional target journal. Adjusts severity calibration and the bar for the Executive verdict. Supported: Marketing Science (MKSCI, default), JMR, JCR, Management Science (MS), QJE, AER, ReStud, Econometrica, JF, JFE, RFS. Defaults to MKSCI if not specified.

The Seven Lenses

Each lens runs as a forked subagent (context: fork) so the main conversation stays clean.

Workflow

Phase 0: Pre-flight

1. Resolve manuscript path. If a .pdf, extract text first via pdftotext -layout (or fall back to reading the .tex source under the same project root).
2. Read the manuscript once in the main agent so the synthesizer has a shared mental model.
3. Resolve --journal. Default to MKSCI. Set lens-5 and lens-3 severity thresholds based on the target (top-five econ journals get tighter identification expectations; MKSCI / JMR / JCR / MS get tighter managerial-relevance and substantive-contribution expectations).
4. Create output directory: <project>/seven_pass_<YYYYMMDD>/ for lens reports.

Phase 1: Spawn 7 reviewers in parallel

In a single message, spawn 7 Task tool calls (one per lens). Each subagent gets:

• The manuscript path (to re-read with its own fresh context).
• The lens-specific prompt (rubric summarised below).
• The target journal so it can calibrate severity.
• Instructions to write to <project>/seven_pass_<YYYYMMDD>/lens_<N>_<lens-name>.md.
• Severity tagging: CRITICAL / MAJOR / MINOR.

Lens prompt summaries (each forked subagent receives its lens's rubric plus the manuscript path and target journal):

• Lens 1 (Abstract): Does the first sentence state the question? Does it name the method? Quantify the headline result? State a one-sentence contribution? Do these four things match the body and the table of magnitudes? Marketing abstracts should foreground the managerial / substantive insight, not just the technical contribution.
• Lens 2 (Intro): Does the intro open with a motivating phenomenon (the marketing-paper standard pattern), then research questions, then approach, then numbered contributions, then roadmap? Is the three-stream lit review placed after the hook (not before)? Are findings previewed with magnitudes?
• Lens 3 (Methods / identification): Is every identification assumption stated and defended? Are known violations (selection, omitted variables, reverse causality, measurement error, SUTVA) addressed? For deep-learning / GenAI components (GAN-based features, sparse autoencoders, vision embeddings), are training-set leakage, holdout discipline, and external validity discussed? Is the unit of analysis vs. unit of randomisation / treatment consistent?
• Lens 4 (Results): Does each table read standalone (caption, units, SEs / CIs clarified, clustering documented)? Are magnitudes interpreted in interpretable units (dollars, percentage points, elasticities, lift)? Are units consistent across tables? Are figures legible at 9pt and color-blind safe?
• Lens 5 (Robustness): Does the paper ANTICIPATE the sharp referee's objections (the AE for the target journal)? Are robustness checks motivated, or just listed? For experiments: power, manipulation checks, attention checks. For obs work: placebo, alternative samples, alternative DVs. For text / image ML: alternative embeddings or feature extractors.
• Lens 6 (Prose): Sentences under 30 words on average? Active voice dominant? Hedging proportionate (neither overclaiming nor endless "may suggest")? Paragraph topic sentences? Marketing voice — modest, evidence-anchored, no econ swagger, no psych breathlessness.
• Lens 7 (Citations): For top-15 cited works, does the in-text claim match the cited paper's actual finding direction? Are 2024–2026 contemporaries cited (in vision-ML / GenAI-marketing, two-year-old citations are a red flag)? Are the marketing canon AND the relevant CS / methods canon both represented?

Phase 2: Synthesize

Wait for all 7 lens reports. Read them and produce <project>/seven_pass_<YYYYMMDD>.md:

# Seven-Pass Review: <Manuscript>

**Date:** YYYY-MM-DD
**Path:** <manuscript>
**Target journal:** <MKSCI / JMR / JCR / MS / QJE / ...>

## Executive verdict

**Quality score:** <NN>/100
**Threshold reached:** <80 desk-survives / 90 R&R-likely / 95 conditional-accept range>
**Overall state:** <SUBMIT / REVISE-MINOR / REVISE-MAJOR / REJECT-AND-RESTART>

Quality threshold key:
- **>=95** Conditional-accept range at the target journal (rare).
- **>=90** Likely R&R with revisable issues.
- **>=80** Desk-survives but needs substantive work before submission.
- **<80** Not yet submission-ready.

## Cross-lens CRITICAL issues
| # | Lens(es) | Issue | Recommendation |
|---|---|---|---|

## MAJOR issues (second-round)
| # | Lens(es) | Issue |
|---|---|---|

## MINOR polish
[bulleted]

## Per-lens scorecard
| Lens | Critical | Major | Minor | Score/10 |
|---|---|---|---|---|
| 1. Abstract | | | | |
| 2. Intro | | | | |
| 3. Methods | | | | |
| 4. Results | | | | |
| 5. Robustness | | | | |
| 6. Prose | | | | |
| 7. Citations | | | | |
| **Overall** | | | | |

## Revision plan (in recommended order)
1. <Highest-leverage fix — usually a lens with 2+ CRITICALs>
2. ...
7. <Lowest-leverage polish>

## Contradictions between lenses
[If two lenses disagree, surface here. E.g., Lens 2 says "expand contribution framing" but Lens 6 says "trim intro by 15%". Recommend a reconciliation.]

The synthesizer must convert the seven lens-level severity counts into the overall quality score:

• Start at 100.
• Subtract 8 per CRITICAL, 3 per MAJOR, 1 per MINOR.
• Floor at 40 (below that, the paper is REJECT-AND-RESTART regardless of arithmetic).

Phase 3: Token-budget report

After synthesis, print:

Seven-pass review complete.
Subagents: 7 (parallel) + 1 synthesizer.
Approx token usage: ~80-120k (vs ~10-15k for /review-paper-light, ~25-35k for /review-paper).
Runtime: ~3-5 min wall-clock.
For cheaper alternatives:
  - 2-agent fast pass: /review-paper-light
  - 6-agent journal-targeted referee report: /review-paper

When to use this skill

• Before first submission to MKSCI / JMR / JCR / MS / QJE / AER / RFS / JF.
• After a major revision when drift is likely (e.g., R1 -> R2 turnaround).
• R&R when referees disagree — surfaces contradictions the revision must navigate.
• Job-market-paper polish where every lens matters.

When NOT to use

• Very early drafts (use /review-paper-light for a fast contribution / identification gut-check).
• Short notes, comments, replies (overkill).
• Already-reviewed papers where nothing substantive changed since the last seven-pass run.

Failure modes

• Manuscript unreadable or empty. .pdf text extraction failed; .tex has unresolved \input{}s. Fall back to the most recent .tex source under the project root; if still empty, abort and surface a clear error.
• Subagent timeout or empty lens report. Re-spawn just the affected lens once; if it fails again, mark that lens as "DEGRADED" in the synthesis and continue. Do not block the report on one lens.
• Lens 7 cannot resolve citations. If the .bib is missing or /validate-bib / Zotero MCP is unavailable, scope Lens 7 to in-text-claim spot checks only and flag in the synthesis.
• Project root not under Overleaf path. Skill writes lens reports next to the manuscript path the user provided; if that directory is read-only, fall back to ~/.claude/cache/seven_pass_<stem>_<date>/ and surface the path.
• No journal specified and project ambiguous. Default to MKSCI; surface this in the report header so the user can re-run with --journal JMR if they want a different calibration.

Out of scope

• Auto-applying fixes. This skill produces a checklist; it does not edit the manuscript. For surgical edits, use /draft or a manual rewrite pass.
• Numeric reproducibility. That is /audit-reproducibility. Lens 4 only flags qualitative inconsistencies (e.g., "Table 3 caption says percentage points but the text reads as proportions").
• Replication-package audit. Use /review-paper-code for that.
• Human judgment. A trusted senior reviewer who knows the subfield still beats seven LLMs. Treat this as triage.

Cross-references

• /review-paper — six-agent journal-targeted referee report (cheaper, faster).
• /review-paper-light — two-agent fast contribution / identification check (cheapest).
• /review-paper-code — code + paper reproducibility audit.
• /audit-reproducibility — numeric-claim-by-claim verification.
• /review-pap — preregistration-stage analogue.