De AI Academic Prose

De-AI Academic Prose

Overview

Use this skill when academic text is technically correct but sounds overly generated, templated, or mechanically balanced. The goal is to preserve the claims, evidence, and formal register while reducing obvious AI-style prose patterns.

This skill also supports whole-document triage for Markdown and LaTeX manuscripts when the user wants structural statistics before rewriting.

Read references/markers.md when you need a concrete checklist of AI-style markers, diagnostic tags, severity levels, or output mode guidance.

Read references/genres.md when the passage belongs to a specific genre such as an abstract, introduction, literature review, methodology, discussion, conclusion, proposal, or statement.

Read references/variance.md when the user wants the rewrite to consider whether the prose needs more authored irregularity, variation, or anti-detection texture.

For long documents or .tex inputs, use scripts/analyze_academic_prose.py to:

• extract section content from LaTeX into normalized Markdown;
• compute whole-document and per-section word-frequency statistics;
• compute section priority and marker density for rewrite planning;
• generate a rewrite queue ordered by likely payoff;
• rank candidate AI-heavy sentences before rewriting;
• save a reusable Markdown or JSON report when the user wants artifacts.

Workflow

-1. Confirm screening parameters before analysis

Before running the analysis or assigning diagnostic tags, confirm the user's preferences for de-AI work. If the user already stated these preferences in the conversation, do not ask again; carry them into the next steps.

At minimum confirm or infer:

• scope: whole document, selected sections, or sentence-level cleanup;
• rewrite pressure: light, medium, or aggressive;
• tolerance for signposted academic English;
• soft discourse marker cap per paragraph;
• variance setting: off, light, or moderate;
• whether the user wants only diagnosis, a rewrite queue, or direct edits.

Keep this preference intake short. In Default mode, ask concise plain-text questions only when the parameters are not already clear and guessing would materially change the outcome.

Do not move on to rewriting when the user's intended meaning is still ambiguous in a way that would materially affect style. If "remove AI tone" could plausibly mean either light cleanup or aggressive de-templating, ask first and wait for the answer instead of silently picking one.

These are blocking parameters for direct edits when unclear:

• how much signposting or connective language the user wants to keep;
• whether the user prefers authored smoothness or more compressed prose;
• whether the user wants diagnosis first or immediate rewriting.

Default assumptions when the user does not specify:

• rewrite pressure: medium
• variance: off
• signposted-English tolerance: preserve comfortable signposting
• soft discourse marker cap: 2 per paragraph
• mode: diagnose-only until rewrite preferences are clear

0. Start with data analysis

Default to document analysis before diagnosis whenever the input is more than a short excerpt. Use the confirmed preference profile from Step -1 to set the analysis thresholds before you run the script or begin mental triage.

Use the script before diagnosis or rewriting when the input is:

• a whole report, dissertation chapter, or proposal;
• a .tex source file;
• a Markdown manuscript with multiple sections;
• a request for statistics, hotspots, or "which parts sound AI-written?"

Typical command pattern:

uv run python scripts/analyze_academic_prose.py /path/to/file.tex \
  --emit-markdown /tmp/extracted.md \
  --emit-report /tmp/analysis.md

If the environment does not use uv, run the script with the local Python launcher required by that environment.

Use the report to decide:

• whether the problem is local or document-wide;
• which sections have the densest template language;
• which sections should be rewritten first;
• which sentences are best candidates for targeted rewriting;
• whether the section balance or sentence cadence looks suspiciously uniform.

Treat soft discourse markers more leniently than stronger template signals. This includes words and phrases such as however, moreover, overall, therefore, furthermore, additionally, consequently, thus, nevertheless, nonetheless, in addition, and on the other hand. Do not flag them just because they appear once. Instead, use the script's paragraph-level cap and only treat them as a style problem when a paragraph exceeds the configured limit. The default cap is two soft discourse markers per paragraph, unless the user prefers an even more signposted style.

For short excerpts, still do a lightweight analysis mentally before assigning diagnostic tags: note repeated phrases, sentence-shape uniformity, and whether the paragraph ends in generic recap language.

1. Diagnose after the analysis pass

Use the analysis results to decide whether the problem is actually "AI tone" rather than weak content, poor structure, or missing evidence.

Assign 2-4 diagnostic tags before you rewrite. Reuse the tag names from references/markers.md. Typical tags include:

• repetitive framing
• meta-writing
• template transitions
• abstract drift
• flat cadence
• generic recap
• over-balanced contrast

Cross-sentence tags (read markers.md for definitions and examples):

• broken connector logic
• pseudo progression
• circular paragraph
• forced symmetry
• weak evidence link
• paragraph drift

The cross-sentence tags cannot be detected by the analysis script. The script reports density and uniformity of local markers, not whether the logic between sentences actually holds. To assign any of the cross-sentence tags you must read the passage yourself and judge the relationship between adjacent sentences or across a whole paragraph. When you do assign one, name the specific sentence or sentence pair where the broken relationship appears — not just the paragraph.

Then assign a severity level based on both the local prose and the document statistics:

• low: locally templated, but the passage mostly sounds authored;
• medium: the generated feel is noticeable across multiple sentences;
• high: the paragraph or section reads persistently formulaic or synthetic.

Focus on passages with:

• repeated framing phrases such as this project, the present project, this framework, this means, therefore, however;
• excessive meta-writing about the purpose of the text rather than direct argument;
• repetitive not X but Y or rather than constructions;
• uniformly balanced paragraph rhythm with low variation in sentence length;
• abstract nouns replacing concrete claims or specific objects;
• low-information summary sentences at the ends of paragraphs.

For this whole class of discourse markers, only escalate when the paragraph frequency is high relative to the configured cap. A single well-placed use is usually acceptable in academic prose, especially when the author naturally prefers a more signposted English style.

2. Preserve the non-negotiables

Do not change:

• technical meaning;
• mathematical notation or symbol definitions unless the user asks for it;
• citations, unless they are plainly wrong or missing;
• the strength of the claim, especially where hedging is deliberate;
• discipline-appropriate academic tone.

The task is not to make the writing casual. The task is to make it sound like a real author wrote it.

3. Identify the genre before choosing rewrite pressure

Determine the genre or closest local function of the passage. Use references/genres.md when the answer is not obvious.

At minimum distinguish among:

• abstract;
• introduction;
• literature review;
• methodology or modelling section;
• results;
• discussion or conclusion;
• proposal, statement, or application materials.

Adjust rewrite pressure to the genre:

• preserve standard phrasing more aggressively in methods and notation-heavy sections;
• preserve distinctions among cited works in literature reviews;
• keep scope, contribution claims, and evidence boundaries tight in abstracts, discussions, and conclusions;
• reduce polished-template rhetoric more aggressively in proposals and statements, while keeping formal tone.

4. Decide whether to add controlled variance

Treat variance as optional. Only apply it when the user asks for it directly, when anti-detection concerns are explicit, or when the passage still feels too uniform after a normal rewrite.

Use one of these settings:

• off: no deliberate irregularity beyond normal good writing;
• light: small cadence variation and a few less-predictable phrasings;
• moderate: more visible sentence-shape variation and stronger reduction of smooth template flow.

Do not use a stronger setting unless the user explicitly asks for it. Follow references/variance.md for allowed and forbidden forms of variance.

When variance is enabled, prefer concrete anti-template moves over vague "randomness":

• break overly even sentence rhythm;
• convert some noun-heavy lists into more natural phrasing;
• allow an occasional brief explanatory bridge sentence when it clarifies why a comparison or modelling choice matters;
• reduce perfectly symmetric conclusion patterns when the evidence does not require that symmetry.

5. Rewrite toward specificity

Prefer:

• concrete subjects over meta subjects;
• direct claims over self-explanation;
• specific system conditions, variables, constraints, and results over abstract umbrella nouns;
• varied sentence lengths and paragraph cadence;
• enough transition words to keep the prose comfortable to read.

Useful rewrite moves:

• Replace This project / the present study with the actual subject.
• Replace This framework allows... with a direct statement of what the model, method, or result does.
• Delete throat-clearing phrases such as it is important to note that.
• Delete paragraph-ending recap sentences that add no new information.
• Merge or split sentences to avoid a flat rhythm.
• Keep helpful connective language when it genuinely guides the reader.
• Remove only redundant, stacked, or overly formulaic transitions.

Do not optimize for the fewest possible connectives. The goal is natural, authored readability, not stripped-down minimalism.

Match the intensity of the rewrite to the severity:

• low: make local edits, keep most sentence structure, and remove only the clearest markers;
• medium: rewrite sentences more directly and vary cadence where multiple markers interact;
• high: rebuild the paragraph around its actual claims, while preserving all evidence, notation, and citation logic.

If variance is enabled, introduce it through rhythm, clause structure, information order, and selective lexical variation, not through invented content or fake specificity.

6. Keep the rewrite grounded

Apply genre-specific guardrails from references/genres.md. When in doubt, be conservative in methods and more interventionist in expository prose around the methods.

Output Pattern

Pick the output mode that best matches the user request. If the user does not specify one, default to diagnose + rewrite.

Supported modes:

1. rewrite-only
2. diagnose-only
3. diagnose + rewrite
4. before/after + rationale
5. minimal edit
6. aggressive rewrite
7. line-by-line edit notes
8. document triage + rewrite plan

Mode guidance:

• rewrite-only: return only the revised passage.
• diagnose-only: report the genre, severity, and strongest diagnostic tags without rewriting.
• diagnose + rewrite: briefly state genre, severity, and 2-4 tags, then rewrite.
• before/after + rationale: show original and revision, then briefly explain the main changes.
• minimal edit: keep sentence structure as intact as possible while removing high-signal markers.
• aggressive rewrite: rebuild the prose more substantially, but keep claims, support, and tone stable.
• line-by-line edit notes: useful when the user wants to learn the pattern of edits rather than just receive a polished rewrite.
• document triage + rewrite plan: run the file analysis workflow first, then report section stats, recurring markers, and the highest-priority passages to rewrite.

Optional controls:

• variance: off | light | moderate
• If not specified, default to off.
• When variance is enabled, briefly state whether it was applied and at what level unless the user asks for rewrite-only output.

Unless the user asks for rewrite-only output, include:

1. analysis summary first;
2. genre;
3. severity;
4. 2-4 diagnostic tags;
5. variance setting when used;
6. the rewrite or analysis requested by the chosen mode.

The analysis summary should lead with:

1. genre;
2. file type (.md or .tex) when relevant;
3. total word count and main-section balance for long documents;
4. section priority, marker density, and the top of the rewrite queue;
5. the most frequent content words or phrases;
6. the top candidate AI-heavy sentences and why they were flagged.

When the script is used, also include:

1. the path to the generated Markdown extraction when one was produced;
2. the path to the generated report when one was produced.

When editing files directly, prefer local rewrites over whole-section replacement unless the section is uniformly affected.

Red Lines

• Do not introduce conversational or journalistic style.
• Do not add motivational filler or rhetorical flourish.
• Do not invent examples, results, or literature support.
• Do not remove necessary methodological caution.
• Do not simulate "human randomness" by adding errors, false citations, or unsupported detail.