Latex Prose Audit

LaTeX Prose Audit

Core principle

Audit the manuscript's sentence-level contract before polishing its surface. First identify who or what performs an action, what the verb actually claims, and what noun phrase the reader is asked to interpret. Then check terminology, articles, modifiers, sentence attachment, captions, equations, and LaTeX syntax. Do not turn one author's wording into a universal prohibition.

Use this skill when

• the target is one or more .tex files, an included LaTeX file, or a LaTeX-derived manuscript;
• the user asks to audit, proofread, revise, or fix academic writing and formatting;
• the paper contains equations, algorithms, tables, figures, citations, a supplement, or quantitative claims that must remain aligned.

Do not use it for a syntax-only compilation failure with no prose or semantic question; use the project's LaTeX diagnostic workflow for that narrower task.

Required reading and scope

1. Read the target file and every relevant \input, \include, abstract, conclusion, supplement, and bibliography entry. Build a file/section map before reporting line numbers. Enumerate sibling submission artifacts such as highlights, a graphical abstract, cover letter, and response letter even when the main file does not include them.
2. Inspect the document class, loaded packages, custom macros, citation commands, and local formatting conventions. Apply IEEE rules only when the target template actually uses them.
3. If the user supplies reference papers, read the complete relevant sections, not only titles or abstracts. Read reference-prose-analysis.md when the target uses the calibrated MEC/CMOEA vocabulary or result-writing patterns. Use it to retrieve fine-grained examples, not to import a paper's overall organization.
4. Read rules-and-checks.md before the first full audit. Use its rules as checks with evidence, not as a word blacklist.

Audit workflow

0. Establish the artifact boundary

Record the files, included children, language, document class, citation system, and requested mode. Preserve the user's scope: audit only unless the user authorizes edits or selects --fix.

1. Build four small ledgers

• Vocabulary ledger: term, meaning, grammatical role, abbreviation, and approved variants.
• Notation ledger: symbol, quantity, domain/index, first definition, and every cross-file use.
• Claim ledger: exact claim, artifact or citation supporting it, claim strength, aggregation/statistical unit, and remaining limitation.
• Artifact ledger: equation/algorithm/figure/table/paragraph references and whether prose, math, pseudocode, and captions describe the same object.

Do not invent a canonical term when the paper has not yet defined one; report the missing decision instead.

2. Run the four rule categories in order

1. Category A — semantic roles and reference: identify the grammatical subject and classify it as an intentional actor, a neutral reporting artifact, or an inanimate object. Only an algorithm, controller, system, author, or explicitly modeled agent performs decisions or claims results. The complete Rule 1 non-agent list covers algorithm parts; evidence objects including metrics, indicators, tests, values, results, data, trends, tables, figures, curves, equations, sections, and experiments; and mathematical or physical objects including objectives, constraints, formulations, representations, populations, archives, states, paths, and costs. Then check canonical terminology, abbreviations, the/a/an, zero article, and evidence verb strength.
2. Category B — sentence-level English: check direct verb choice, collocations, modifiers, numerical qualifiers, pronouns, which/where, parentheticals, respectively, tense, voice, parallelism, and repetition.
3. Category C — mathematical and artifact prose: check scalar/vector/ matrix/set/function notation, indices and superscripts, domains, ordered sequences, dimensions, units, formula semantics, algorithm line prose, equations, figures, tables, captions, ranges, and numerical typography.
4. Category D — LaTeX source integrity: check references, labels, citation commands, spacing, equation punctuation, protected commands, math/text boundaries, URLs, custom macros, and compilation/render verification.

Revise naturalness only after these four categories; preserve the author's technical meaning and language.

3. Classify findings

• Blocker: compilation break, undefined or contradictory notation, missing algorithm step, invalid comparison, or an unsupported guarantee that changes the scientific meaning.
• High: claim-evidence mismatch, cross-document inconsistency, wrong constraint/metric semantics, or terminology drift affecting interpretation.
• Medium: unclear result ordering, causal overstatement, paragraph overload, ambiguous reference, or repeated definition.
• Low: local awkwardness, caption polish, punctuation, or template style.

Use --quick only for Blocker and High findings. The default full audit reports all actionable findings. --fix permits minimal edits after preserving a before/after mapping; it does not authorize redesigning the model, replacing citations, or silently changing equations.

4. Produce the audit

Lead with the verdict and highest-risk findings. Use this table, sorted by priority:

Quote only the smallest necessary phrase. Give a replacement sentence when it is safe; otherwise state what must be verified and why an automatic edit would be unsafe. Finish with 2–3 positive patterns, an unresolved-check list, and a short count by priority.

5. Verify edits

After --fix or an explicitly authorized revision:

1. re-read every changed paragraph and its neighboring equation, figure, citation, or algorithm reference;
2. rerun the ledgers and search for stale terms, labels, and symbols;
3. run the project's LaTeX compiler (normally latexmk with the target's engine) when available, and report the exact command and result;
4. run git diff --check when the target is in a Git working tree.

If compilation is unavailable or the source is incomplete, report that status explicitly. Never claim that a LaTeX fix is verified from a static text scan alone.

Boundaries

• Do not rewrite bibliography records or manufacture citations.
• Do not treat a reference paper's style as proof that its wording is correct.
• Treat every algorithm part listed in Rule 1 as a structural object, not an independent actor. When one is the subject of an executed action, change the subject to the algorithm/controller or use passive voice. Retain only structural descriptions such as consists of, comprises, contains, is designed to, and is applied to.
• Do not enforce one citation syntax, article pattern, math delimiter, or non-breaking-space convention without checking the target template.
• Do not edit comments, verbatim blocks, labels, citation keys, macro names, or mathematical expressions as if they were ordinary prose.