Execute a comprehensive, multi-pass diagnostic audit of an academic or technical manuscript, producing a structured improvement report that identifies issues across 24 audit dimensions — from macro-coherence and argumentative architecture through claims-evidence calibration, narrative flow, prose microstructure, rendered visual inspection, and cross-element coherence, down to citation hygiene and reproducibility.

The output is a prioritized, actionable improvement plan — not a line edit. The goal is to surface structural, logical, and clarity issues that authors systematically miss because they're too close to the text.

Optimized for arXiv/preprint submissions with flexible compliance standards.

Companion skill: audits whether manuscript content (numbers, tables, figures, ordering, terminology) is computationally derived from code and scripts. This skill audits the document as prose; that skill audits computational grounding. Run both for complete pre-publication coverage.

Rule: This skill never opens the codebase. manuscript-provenance never judges prose quality. Each reads the other's report when available.

Integration point — Macro Manifest: manuscript-provenance produces a macro manifest as part of its §4 audit: a structured list of every macro-injected value, its resolved numeric value, its source (script + output file), and its location(s) in the manuscript text. This skill's Pass 13 (Cross-Element Coherence) consumes that manifest to check whether the prose surrounding each injected value is appropriate for the actual value. If no provenance report exists, this skill extracts macro values directly from source (less precise — no source tracing, but coherence check still runs).

Read the uploaded manuscript. Accept PDF, DOCX, LaTeX source, or Markdown. If multiple files are uploaded (e.g., main text + supplementary), process all of them.

Identify:

For arXiv submissions, compliance checks are advisory. Focus on technical quality, reproducibility, and clarity rather than strict formatting rules.

Read — the comprehensive 24-section, ~175-checkpoint refactoring checklist. Every audit pass is structured against this checklist.

Read references/checklist.md

Execute the following passes sequentially. Each pass maps to one or more checklist sections. Work systematically — for each checkpoint:

Scan prose sections for residual AI-writing patterns using detection rules from . Academic manuscripts drafted or polished with AI assistants often retain detectable tells.

Focus on patterns relevant to academic writing:

Skip patterns that are acceptable in academic prose:

This pass is MEDIUM priority. Flag findings but do not over-correct — academic conventions overlap with some AI patterns. Severity: report individual instances as LOW, flag clusters of 3+ patterns in a single paragraph as MEDIUM.

This is a dedicated pass through every assertion in the manuscript.

For each claim:

This pass is HIGH priority. Claims-evidence mismatch is the single most common reason reviewers reject papers. An overclaim in the abstract poisons the entire reading.

Read the manuscript linearly, tracking the reader's cognitive state. At each sentence and paragraph boundary, check:

Flag any location where a domain-expert reader would need to re-read, scroll back, or pause to reconstruct the logical connection. These are flow breaks.

This pass is HIGH priority. Papers with strong results but poor narrative flow exhaust reviewers. A reader who has to fight the text stops trusting the author.

Sentence-level and paragraph-level patterns that compound into readability problems:

This pass is MEDIUM priority on individual items but compounds — a manuscript with 20 ambiguous pronouns, 10 density spikes, and 5 dangling modifiers is materially harder to read even though no single instance is fatal.

This pass requires the compiled PDF. If only LaTeX source is provided, ask the user for the compiled PDF or compile it.

Open the PDF and inspect every page at actual print scale:

This pass is HIGH priority. A paper with illegible axis labels or a table split across pages signals carelessness to reviewers regardless of technical quality. These defects are invisible from source and the author often doesn't notice because they read the paper in their editor, not in the compiled output.

Read the manuscript as an integrated system. For each figure, table, and macro-injected value:

If a manuscript-provenance report exists, load its macro manifest (list of all traced macro values with locations and source values) and use it as input for step 4. If no provenance report exists, extract macro values directly from source.

This pass is HIGH priority. Cross-element incoherence is the most insidious class of manuscript defect — each piece looks fine in isolation, the system is broken. Reviewers notice because they read the document linearly and encounter contradictions the author can't see because they edit pieces independently.

Note for arXiv: Ethics statements, anonymization, page limits, and strict formatting requirements are marked N/A by default. Focus on technical quality, reproducibility, and clarity.

Produce the report as a structured document. Use as the output format.

Read references/report-template.md

After completing the full scan, categorize issues:

For arXiv submissions, focus HIGH priority on technical quality and reproducibility. Compliance items (ethics statements, formatting) are typically LOW priority or N/A.

Present the priority queue first, then the detailed findings.

Save the report as a Markdown file in the same directory as the manuscript, named .

Present the file to the user with a concise summary:

User says any of:

All trigger this skill. Partial reviews (e.g., "just check citations") still run the full audit — the user benefits from comprehensive diagnostics even when they only asked about one aspect.