academic-writing

1. Anti-Pattern Detection

Five diagnostic patterns that mark AI-generated research prose.

Hedging Soup

Stacking uncertainty markers drains every sentence of meaning.

Diagnostic: count hedging words (potentially, possibly, may, might, seems, could, arguably, perhaps) per paragraph. More than two in a single paragraph signals hedging soup. Replace stacked hedges with one precise statement of what is known and one precise statement of what limits that knowledge.

Formulaic Transitions

"Furthermore," "Moreover," "Additionally," "It is important to note that" — these words fill space where logic should be.

Diagnostic: scan for Furthermore/Moreover/Additionally/It is important to note/It is worth mentioning/Interestingly. If any appear, ask whether the logical relationship is already clear from sentence order. If yes, delete the transition. If no, replace with the actual relationship (but, because, so, despite, after).

Structural Monotony

Every paragraph follows the same template: topic sentence, three supporting points of equal weight, summary sentence. Real academic writing varies its rhythm.

Diagnostic: map paragraph lengths across a section. If more than three consecutive paragraphs fall within 10 words of each other, restructure. Vary paragraph length by at least 30%. Mix single-claim paragraphs with multi-sentence analytical passages.

Abstraction Fog

Categories and generalities where specifics belong.

Diagnostic: search for "various," "different," "multiple," "numerous," "important," "significant" (non-statistical), "the field," "the literature." Each one is a prompt to substitute a specific name, number, or concrete referent.

Voice Erasure

The writer disappears behind passive constructions and impersonal phrases, even when the discipline expects authorial presence.

Diagnostic: count instances of "it can be," "it is suggested," "it was found," "it should be noted." Replace each with the actual agent performing the action. In the sciences and social sciences, first person plural is standard. In the humanities, first person singular. Know the convention for the target discipline.

2. Discipline-Aware Register

Academic writing conventions vary by field. Adjust register to the target discipline.

• Sciences (STEM): Concise, method-focused, passive voice acceptable for methods sections, active voice for argumentation. Numbered hypotheses. Statistical reporting follows APA or field-specific standards.
• Social Sciences: Balance of active and passive voice. First person plural standard. Theoretical framing expected in introductions. Effect sizes alongside p-values.
• Humanities: First person singular common. Longer analytical paragraphs. Close reading and interpretive argument valued. Direct engagement with other scholars' positions.
• Interdisciplinary: Default to active voice with first person plural. Define terms from each contributing field. Bridge jargon gaps explicitly.

3. Citation Integration

How sources enter sentences matters as much as which sources you cite.

• Narrative citation when the author's identity matters: "Foucault (1975) argued that..."
• Parenthetical citation when the finding matters more than who found it: "Incarceration rates tripled between 1980 and 2000 (Alexander, 2010)."
• Direct quotation only when the original wording is the point: a key definition, a contested phrase, a striking formulation. Never quote to avoid paraphrasing.
• Synthesis citation to show consensus or disagreement: "Several studies confirm this pattern (Lee, 2019; Nakamura, 2020; dos Santos, 2021), though one found no effect in adolescents (Byrne, 2022)."

When producing any research prose, follow this sequence:

1. Identify the discipline and audience. Adjust register, citation style, and voice conventions accordingly.
2. Draft with specificity. Name studies, methods, sample sizes, years. Replace every "various studies" and "the literature suggests" with concrete referents during drafting, not after.
3. Vary structure deliberately. Before writing each paragraph, choose its shape: single-claim, multi-evidence analytical, pivot-to-complication, or narrative-of-debate. Do not repeat the same shape for three consecutive paragraphs.
4. Use transitions only for actual logical relationships. If deleting a transition word doesn't change the meaning, delete it.
5. Self-audit before presenting. Run these five checks against the completed draft:

Self-Audit Checklist

• Hedging: Fewer than two hedging words per paragraph. Every uncertainty is expressed through evidence ("3 of 5 studies") not through stacked qualifiers.
• Transitions: Zero instances of Furthermore/Moreover/Additionally/It is important to note. Every remaining transition expresses contrast, cause, consequence, or concession.
• Structure: No three consecutive paragraphs within 10 words of each other in length. At least two different paragraph shapes per section.
• Specificity: Zero instances of "various studies," "the literature," "multiple contexts," "important implications" without a concrete referent following within the same sentence.
• Voice: Fewer than three instances of "it can be," "it is suggested," "it was found" per page. First person used where discipline permits.
• Citations verified: If the draft cites scholarly sources, write them to a file (one per line or JSON) and run uv run <literature-review-dir>/scripts/verify_citations.py --input refs.txt. Exit 0 required — any mismatched or not_found citation is corrected or removed before presenting. Never present prose containing a citation the verifier could not resolve.
• Claims traced: If the draft attributes findings or context to scholarly sources present in a review workspace, including abstract-only sources, save the completed draft at $DRAFT and write a non-empty claims.json. Every evidence or background entry contains claim, paper_id, citation (one Author, year or [n] identity), and a verbatim supporting_quote; a claim citing two sources needs two entries. Confirm corpus.json contains trusted authors, year, and role metadata. Run uv run <literature-review-dir>/scripts/check_claims.py --claims claims.json --workspace "$WS" --synthesis "$DRAFT". For numeric citations, first verify the ordered bibliography and add --references refs.json. Exit 0 is required. Resolve every needs_review by reading the reported span. An invalid_binding means the citation did not resolve to exactly one trusted corpus record; use its reason_code to fix metadata or the citation, rather than retrying quote matching. Background entries may cover number-free context only, and every non-year number in a cited synthesis sentence must appear in verified matching claims for each cited identity.

If any check fails, revise before presenting the output.