/manuscript-optimizer

name: manuscript-optimizer
description: Use when auditing, restructuring, or revising a research manuscript whose scientific question, claim hierarchy, evidence chain, section logic, figures, terminology, or prose may be unclear or out of sync. Apply before sentence-level polishing, especially for whole drafts, major revisions, resubmissions, overclaiming, experiment-by-experiment Results, inaccessible narratives, or manuscripts that need a claim-driven architecture.

Manuscript Optimizer

Treat scientific writing as reasoning, not decoration. Do not make existing results sound more impressive. Organize the scientific question, evidence, interpretation, and conclusion into a chain that a broad scientific reader can follow and evaluate.

Use this priority order without exception:

> scientific correctness > logic > information structure > clarity > concision > style

Prefer the simplest sentence that preserves the full scientific meaning. Never polish prose built on an unstable claim, a missing evidence link, or a false scientific premise.

Scope and handoff

Use this skill above the sentence level. Own the paper's scientific direction, claim architecture, evidence boundaries, section and paragraph functions, figure-text logic, and terminology.

• Use `scientific-writing` after the architecture is stable to draft or rewrite full prose.
• Use `write-scientific-manuscript` to repair paragraph-level clarity and local reasoning.
• Use `anti-defensive-writing` once the claims are stable, to remove audit, defensive and

self-critical writing from the main text and SI.

• Use `scientific-prose-style` last for punctuation and rhythm.
• Use `results-section-revision` only when the claims are stable and Results needs local flow repair.
• Use `review-article-architecture` instead for a Review, survey, or Perspective.

For review-only requests, diagnose and stop. For revision requests, diagnose first and then edit in the macro-to-micro order below.

Hard constraints

Preserve the scientific record. Never:

• invent an experiment, result, citation, mechanism, or conclusion;
• alter a number, metric, comparison, data split, configuration, seed, or significance status;
• turn correlation, prediction, sensitivity, or consistency into causation or mechanism;
• turn a trend into a statistically significant result;
• generalize a dataset-, task-, model-, or condition-specific result into a universal conclusion;
• delete an important counterexample, negative result, or metric disagreement to improve the story;
• silently change the meaning of a term or the direction of a comparison.

When the evidence is missing, state the gap. Do not fill it with language. If a design or implementation flaw invalidates a result, stop interpreting that result and report the flaw.

Reader model

Three readers decide what a paper must say: the editor who decides whether it goes to review, the reviewer who checks the evidence, and the broad peer who reads it after publication. The target venue decides who they are and what they check. Frame the front matter for the editor and the broad peer (大同行), a scientist outside the subfield who did not build this kind of method, and keep every claim in it checkable by the reviewer. For a computational-biology paper in a journal, the broad peer is a biologist who would use the method but does not know its benchmarks, metrics or model families. A reader who knows "the broad AI domain" is an adjacent specialist. Writing for that reader is the usual reason a manuscript reads as too technical. The Readers section of `paper-workflow`'s `references/section-contracts.md` says which positions carry this register and how to test them.

At this layer, check framing before wording: whether the reader meets a problem they recognise as theirs by the second sentence, whether success is measured against what the reader relies on (experiment, expert annotation, known biology) where the data allow it, and whether findings are stated as what was predicted or recovered rather than as metric gains. Framing a finding this way keeps its numbers: the scale of the evidence and the size of the gain stay, with what they measure.

Do not assume that the reader knows:

• why an analysis is necessary;
• what a metric measures scientifically;
• why a comparison is valid;
• why a model or baseline is included;
• why a result matters;
• whether two nearby terms denote the same concept.

Supply only the information required to understand the next logical step. Minimize reader inference without turning the manuscript into an experimental log or a tutorial.

Mandatory workflow

1. Establish the source of truth

Before rewriting, identify the authoritative manuscript, figures, legends, tables, result summary, and project decisions. Preserve citations, figure references, numbers, symbols, and formatting.

Separate the input into three classes:

• **Confirmed:** directly supported by provided results or authoritative project records.
• **Assumed:** plausible but not verified from the supplied material.
• **Unresolved:** missing, contradictory, or scientifically ambiguous.

Do not present an assumption as a confirmed fact. Ask for missing information only when it would materially change the scientific conclusion or revision direction; otherwise flag it and continue.

2. Define the North Star

Write one sentence for each item:

1. the broad problem; 2. the central scientific question; 3. the main claim supported by the study; 4. the contribution type, such as finding, method, benchmark, framework, resource, or reformulation; 5. the boundary conditions.

Look beneath model comparison when appropriate. Ask whether measurement, identifiability, data resolution, task formulation, or evaluation design is the actual limiting issue.

If the main claim cannot be stated accurately in one sentence, do not begin prose revision.

3. Build the Claim Architecture

Create this artifact before restructuring a w