Manuscript review revision

Manuscript Review & Revision

Operating contract

Treat review, revision, and formatting as separate phases. Use this non-negotiable order:

target journal -> journal profile -> frozen input -> independent review -> synthesis -> user gate -> scientific revision -> reference/figure closure -> language -> formatting -> release gate

• Keep the manuscript read-only through review.
• Do not turn a polished file into a submission-ready claim.
• Do not invent experiments, analyses, citations, journal rules, reviewer identities, line numbers, or completed changes.
• Use PASS, FAIL, or NOT ASSESSABLE for critical gates.
• Explain findings in the user's language while preserving exact manuscript text, filenames, identifiers, and journal wording.

Host capability contract

Load references/platform-compatibility.md before using tools or bundled scripts.

• Resolve the directory containing this SKILL.md as SKILL_ROOT; never assume the current working directory is the Skill directory.
• Map each required behavior to the capabilities available in the current host. Tool names are examples, not requirements.
• For full review, use the host's real non-fork subagent/delegation mechanism with a fresh isolated context for every reviewer.
• Record actual host-provided Agent task IDs, timestamps, frozen-input hashes, and report hashes as the review execution receipt.
• If a required capability is unavailable, mark the affected gate NOT ASSESSABLE; do not imitate a completed capability in one conversation.

Step 0: establish the target journal

Before reviewing, inspect the invocation for an exact target journal.

• If absent, ask only: 本次目标期刊是什么？如果尚未确定，请回复“不确定，请推荐期刊”。 Then stop.
• If an abbreviation is ambiguous, confirm the exact journal name and stop.
• If the exact journal is present, do not ask again.
• If the user is unsure, load references/journal-discovery-and-profile.md, run journal-recommendation, return a verified Top 5, and pause for the user to select a primary target.
• Do not start full review until the primary target is fixed.

For every fixed target, browse current official journal sources. Resolve article type and submission stage from the command or manuscript; ask only when ambiguity would materially change the rules. Record URLs and access dates. If current official sources cannot be inspected, mark journal-specific work NOT ASSESSABLE.

Step 1: freeze and inventory the input

Record:

• manuscript path, modification time, SHA-256, and apparent version
• target journal, article type, and stage
• main text, references, figures, tables, legends, supplements, source data, response letter, and tracked version available
• absent, stale, or mismatched materials

Create 00_input_inventory.json and 01_journal_profile.json. Validate the latter with:

python3 "$SKILL_ROOT/scripts/validate_journal_profile.py" 01_journal_profile.json

Do not silently combine manuscript, figures, or supplements from different versions.

Step 2: select the task mode

Use one or more modes, but preserve their order:

Mode	Purpose
journal-recommendation	Recommend a verified Top 5 when the target is unknown
scientific-review	Default for a supplied manuscript; review only
reference-audit	Verify reference reality, integrity, format, placement, and exact claim support
response-to-reviewers	Triage comments and prepare a traceable response package
revise-manuscript	Revise only after the review pause and explicit authorization
format-manuscript	Apply current official journal requirements after scientific content stabilizes
release-gate	Run final fail-closed submission audit
full-run	Run all phases but still pause after review before any revision

If the user supplies a target and manuscript but no mode, default to scientific-review, read-only.

Step 3: run the independent reviewer panel

Load references/multi-agent-review.md, references/journal-tier-rubrics.md, and references/review-panel-receipt-schema.md, references/concern-ledger-and-adjudication.md completely. Load the applicable sections of references/biomedical-review-gates.md.

• Spawn at least five actual independent reviewer agents with the host's non-fork, isolated subagent/delegation mechanism when available.
• Use five reviewer agents plus a root synthesis; do not count the root synthesis as a reviewer.
• Use exactly five core seats. For high-tier or complex manuscripts, add at most one sixth specialist seat, selected by the highest-risk unresolved trigger. Never add both a figure reviewer and an adversarial reviewer.
• Assign every applicable review axis to exactly one primary owner before dispatch. A reviewer may report outside its owned axes only for a clearly blocking concern.
• Limit each reviewer to eight prioritized concerns and 1,800 word-equivalent units. Completeness means covering the assigned risk surface, not repeating a whole-manuscript review.
• Run agents in waves when concurrency is limited; never reduce the reviewer count to fit one wave.
• Give every reviewer the same frozen manuscript, inventory, factual journal profile, shared fact base, and role-specific rubric.
• Do not expose one reviewer's output to another before all independent reports finish.
• Use functional review lenses, not fabricated human biographies.
• If actual agent delegation is unavailable, do not claim multi-agent review completion. Mark the panel gate NOT ASSESSABLE and ask before using isolated single-agent passes as a fallback.

Create 02_shared_fact_base.md, 03_review_panel_plan.json, and reviews/reviewer_01.md through reviews/reviewer_05.md or higher. The panel plan must use schema 2.1, contain five core seats and no more than one triggered optional seat, record axis ownership and output budgets, and close every reviewer against a unique host task ID, FRESH_NON_FORK context, start/end time, the three frozen input hashes, and the resulting report hash. Validate:

python3 "$SKILL_ROOT/scripts/validate_review_panel.py" 03_review_panel_plan.json

Step 4: synthesize, decide, and pause

After all independent reports are frozen, create and validate reviews/concern_ledger.tsv:

python3 "$SKILL_ROOT/scripts/validate_concern_ledger.py" \
  reviews/concern_ledger.tsv 03_review_panel_plan.json

Then consolidate without averaging away disagreements. Create:

• 04_cross_review_matrix.tsv
• 05_review_verdict.md

Use one matrix row per normalized issue and keep the author-facing verdict at or below 900 word-equivalent units. Put detail in the traceable ledger; do not repeat every reviewer report in the verdict.

Validate the verdict:

python3 "$SKILL_ROOT/scripts/validate_review_verdict.py" 05_review_verdict.md

Use exactly one review posture:

• PROCEED_TO_REVISION
• MAJOR_SCIENTIFIC_REWORK_REQUIRED
• RETARGET_RECOMMENDED
• NOT_ASSESSABLE

Then pause and ask whether the user authorizes revision. Do not edit prose, restyle headings, or create a revised manuscript before this authorization, even during full-run.

Step 5: revise in scientific priority order

After authorization, load references/revision-and-response.md.

Apply:

1. blocking scientific issues
2. major methods, statistics, and validation issues
3. claims, limitations, and references
4. figure-text-legend-data closure
5. article architecture
6. paragraph and sentence language
7. target-journal formatting

Preserve the original. Produce a tracked/review copy, a clean revised copy, and revision_log.tsv. Do not state that a change was made unless it is present in the delivered manuscript.

Step 6: audit references and exact claim support

Load references/reference-integrity.md. Build 06_reference_audit.tsv with one row per atomic claim-citation relationship, then run:

python3 "$SKILL_ROOT/scripts/validate_reference_audit.py" 06_reference_audit.tsv

Metadata verification and semantic support are separate. A real paper can still be the wrong citation. A related title, metadata-only result, or review article does not establish direct support.

Step 7: format as a submission manuscript

Load references/manuscript-formatting.md.

• Treat the official journal template and current stage-specific guidance as authority.
• When no color is explicitly required, use black title, headings, subheadings, and body text.
• Do not use report-style covers, colored heading themes, cards, callouts, banners, icons, decorative rules, or business-document styling in the submission manuscript.
• Keep audit reports separate from the clean manuscript.
• For DOCX, use a reliable document runtime available in the host, run the mechanical style audit, render every page to PNG/PDF, inspect every page, fix, and re-render.

python3 "$SKILL_ROOT/scripts/audit_docx_manuscript_style.py" manuscript.docx

A mechanical pass does not replace official-template or rendered visual review.

Step 8: run the release gate

Load references/release-gates.md. Re-run targeted reviewer agents when revisions changed the claim architecture, core evidence, statistics, figures, or references.

Return exactly one final state:

• RELEASE PASS
• RELEASE FAIL
• RELEASE NOT ASSESSABLE

Never predict acceptance.

Artifact contract

Use this stable order when the corresponding phase runs:

00_input_inventory.json
01_journal_profile.json
02_shared_fact_base.md
03_review_panel_plan.json
reviews/reviewer_01.md ...
reviews/concern_ledger.tsv
04_cross_review_matrix.tsv
05_review_verdict.md
06_reference_audit.tsv
revision/issue_ledger.tsv
revision/manuscript_tracked.docx
revision/manuscript_clean.docx
revision/revision_log.tsv
07_format_audit.json
08_release_gate.md

Keep review artifacts factual and utilitarian. The submission manuscript must not inherit their tables, colors, status boxes, or report styling.

Resource routing

Resource	Load when
references/platform-compatibility.md	Resolving host tools, subagents, bundled scripts, or install-specific behavior
references/journal-discovery-and-profile.md	Target journal is unknown or any journal-specific task begins
references/multi-agent-review.md	Planning, running, or synthesizing reviewer agents
references/journal-tier-rubrics.md	Calibrating reviewer strictness or selecting specialist roles
references/review-panel-receipt-schema.md	Building and validating host task receipts and report hashes
references/concern-ledger-and-adjudication.md	Normalizing findings, evidence anchors, consensus, disagreement, and resolution
references/biomedical-review-gates.md	Applying design-specific clinical, wet-lab, omics, AI, review, or animal stress tests
references/reference-integrity.md	Auditing, adding, moving, or formatting citations
references/revision-and-response.md	Revising a manuscript or responding to reviewers
references/manuscript-formatting.md	Creating or checking DOCX/PDF/LaTeX submission files
references/release-gates.md	Deciding whether a package is ready

Boundaries

• Official instructions, specific editor directions, and supplied templates outrank generalized style guidance.
• A journal benchmark does not authorize imitation or overclaiming.
• Search snippets and metadata do not count as inspected scientific evidence.
• If the user asks only for diagnosis, do not modify files.
• If a scientific direction changes, freeze passed artifacts and reopen review before drafting through the turn.