Academic Writing

Role-Based Exercise Set along the Writing Process

Methods
An exercise that trains the use of generative AI as a writing assistant in multiple roles – with focus on traceability, argument logic, and linguistic precision.
Author
Affiliation

Moritz Mähr

University of Bern

Published

December 24, 2025

Modified

February 12, 2026

Overview and Didactic Goal

This exercise trains the critical use of generative AI as a writing assistant in several clearly delimited roles. The roles are arranged to correspond to a typical academic writing process: clarify problem → develop ideas → structure → draft → revise → finalize.

The goal is not to “automate” text production, but to increase decision quality: better research questions, more consistent arguments, more precise language – while maintaining traceability of authorship and the evidence used.

Prerequisites

  • Basic understanding of historical research methods
  • Basic knowledge of working with generative AI (especially prompting)
NotePrompt Engineering

If you are not yet familiar with prompting, we recommend completing the Prompt Engineering exercise first.

NoteLLM

You can complete this exercise with LLMs from different providers. For this exercise, it is helpful if the LLM has internet access and allows file uploads.

Learning Objectives

  • Treat responses as hypotheses/drafts and systematically verify them.
  • Design prompts so that AI doesn’t “invent” but works with provided material.
  • Draft an argument structure (thesis–evidence–conclusion) and test for consistency.
  • Generate paragraphs from own bullet points and then revise them professionally/terminologically.
  • Conduct revision in two steps: logic/comprehensibility → language/style.
  • Transparently document your AI use (prompt and decision protocol).

Methodological Framework: AI as Writing Infrastructure

In this exercise, AI is understood as a heuristic writing tool: it assists with formulations, structuring, and verification, but it replaces neither source work nor interpretative decisions.(Bender et al. 2021; Campbell 2025; Oberbichler and Petz 2025) Methodologically helpful is a clear division of labor:

  • Historical authorship: Research question, evidence selection, interpretation, normative positions.
  • AI contribution (instrumental): Generate variants, structure suggestions, readability tests, error diagnosis.
  • Verification procedure (mandatory): Every AI statement must either (a) be traceable to your input material or (b) be marked as a hypothesis and externally verified.(The Turing Way Community 2022)

For more reliable work, consistently use two prompt principles: 1. Delimiting the workspace (“work only with my bullet points; mark uncertainties”). 2. Formatted responses (lists/argument maps), so you can systematically verify and selectively adopt.

ImportantBasic Rule: AI May Not “Invent”

Use AI only as a transformation and verification tool for your material (notes, source excerpts, literature summaries, outline drafts). When AI names additional facts, literature, or sources, that counts as a research hypothesis that you must externally verify.

NoteWorking Mode: AI Protocol (Audit Trail)

During the exercise, maintain a brief AI protocol. This is part of scientific traceability (and facilitates later disclosure/methods section).

Minimal Template | Field | Content | |—|—| | Role / Exercise | | | Goal (1 sentence) | | | Input (which material?) | | | Prompt (core excerpt) | | | Response (brief) | | | Verification steps (how verified?) | | | Decision (adopted/adapted/rejected + justification) | |

Prerequisites

  • Own topic fragment (e.g., seminar paper) or a provided text section (approx. ½–1 page)
  • Own bullet points/notes on the topic (at least 10 bullet points)
  • Access to an LLM interface of your choice

Structure of the Exercise

The roles are bundled into phases. Individual steps can be skipped; however, at least one complete round through revision is sensible. Core path (approx. 45–60 min): 1 → 2 → 3 → 4 → 6 → 8.

Optional deepening: 5 (local blockages), 7 (style/translating), 9 (readability test as final review).

  1. Comprehension Checker (clarify task, identify blind spots)
  2. Brainstorming Partner + Devil’s Advocate (open idea space + counter-arguments)
  3. Structurer & Outliner (argument architecture)
  4. Text from Bullet Points (drafting under fact discipline)
  5. Blockage Solver (local writing problems/transitions)
  6. Logic Checker (coherence, evidence, inference rules)
  7. Proofreading + Translator (style/grammar/register; optionally multilingualism)
  8. Abstract Writer (compression, focus, “so what?”)
  9. Comprehension Checker (Re-Run) (readability from target audience perspective)

1. Comprehension Checker

Goal

Make the writing task and your own argumentation intent so precise that later AI steps don’t “evade” or stabilize false premises.

Task

  1. Give AI only the following: Topic in 2–3 sentences, context (course/genre), target audience, scope, submission format.
  2. Have AI ask follow-up questions and mark uncertainties.

Prompt Scaffold

You are a comprehension checker. Ask me 8–12 precise follow-up questions to clarify the writing task and argumentation goal.
Mark:
(1) missing information,
(2) implicit assumptions,
(3) terms that need to be operationally defined.
No content suggestions, only questions and diagnosis.

Context: …
Topic/thesis (preliminary): …
Genre/scope: …
Target audience: …

Critical Comparison

  • Which follow-up questions force you to clarify terms (e.g., “modernization,” “neutrality,” “governance”)?
  • Which questions would you prioritize differently – and why?

Result - 1 revised task (3–5 sentences) + 3 defined key terms

2. Brainstorming Partner + Devil’s Advocate

Goal

Generate ideas without inventing evidence, and immediately co-think counter-arguments/alternative interpretations.

Task A: Brainstorming (idea-oriented, evidence-neutral)

Role: Brainstorming partner. Generate 12 possible argument lines/hypotheses on my topic.
Conditions:
- Formulate each line as a verifiable claim (no facts).
- For each line: which kind of evidence I would need (source types/literature types).
- Avoid specific literature titles or archive signatures; if necessary, formulate as search heuristic.

My topic: …
My material (only types, no sensitive content): …

Task B: Devil’s Advocate (confrontational, but productive)

Role: Devil's Advocate. Choose the 3 most plausible of my hypotheses and argue against each.
For each counter-argument:
- Name the vulnerable assumption.
- Suggest a "stress test" (what finding would overturn my argument?).
- Formulate an alternative interpretation (max. 2 sentences).
No external facts.
Hypotheses: …

Critical Comparison

  • Which hypotheses are falsifiable (in the sense of clear refutation conditions)?
  • Where does presentism threaten (projecting present categories onto the past)?

Result - 1 prioritized hypothesis/thesis + 2 central counter-arguments + 1 planned stress test

3. Structurer & Outliner

Goal

Draft a verifiable argument chain (thesis → sub-theses → evidence → conclusion) before text is generated.

Task

Role: Structurer & Outliner.
Create an outline (max. 8 main points) for a text of approx. X words.
For each point:
- Goal of the section (1 sentence)
- Central claim (1 sentence)
- Required evidence (keywords; "from my notes/sources")
- Transition to next point (1 sentence)

Constraints:
- No new facts.
- If evidence is missing: mark as [EVIDENCE GAP] instead of supplementing.

Material (bullet points/notes): …

Quality Check

  • Does each claim have an evidence trail (at least as placeholder)?
  • Are there sections that only provide “overview” but have no argumentative function?

Result - Outline + marking of 2–3 evidence gaps

4. Text from Bullet Points

Goal

Have paragraphs formulated from your own material without giving up content control.

Task

Select one outline point (e.g., “Main argument 1”) and provide 8–12 bullet points (evidence, observations, quotes/paraphrases, terms).

Role: Text generator from bullet points.
Write a paragraph (120–180 words) in academic English.
Conditions:
- Use only information from my bullet points.
- No new examples, data, names, sources.
- If a sentence would need additional information: mark [UNCERTAIN].
- Keep my terminology; ask if unclear.

Bullet points: …

Critical Comparison

  • Delete all sentences that are not directly traceable to your bullet points.
  • Check terminology: Where does AI smooth terms (e.g., “state,” “society,” “actor”) impermissibly?

Result - 1 paragraph + a version in which you independently refined 3–5 sentences

5. Blockage Solver

Goal

Solve local problems (transitions, opening sentences, term definitions) without having the text “rewritten.”

Micro-Exercises (choose 2)

A) Build Transition

Role: Blockage solver.
Formulate 5 variants for a transition sentence between paragraph A and B.
Variant 1 neutral, 2 more argumentative, 3 more explanatory, 4 shorter, 5 with contrast marker.
Paragraph A: …
Paragraph B: …

B) Operationalize Definition

Formulate 3 possible working definitions for the term "X" in the context of my topic.
For each definition:
- Delimitation from 1 neighboring term
- 1 observable indicator (how do I recognize X in sources?)
No external facts.
Context: …

C) Sentence Diagnosis

Mark in my paragraph:
- the central sentence (topic sentence),
- sentences with unclear referent ("this," "thereby," "one"),
- sentences that actually contain two claims.
Suggest minimal rearrangements (no content rewriting).
Text: …

Result - 1 revised transition/definition + brief justification of why this variant

6. Logic Checker

Goal

Make argumentation logic visible: Do evidence support claims? Are inference rules traceable? Where are premises missing?

Task

Role: Logic checker.
Analyze my text (max. 300 words) as an argument map:
1) Main thesis
2) Sub-theses
3) Explicit evidence (from the text)
4) Implicit premises (that I apparently presuppose)
5) Possible logical jumps or contradictions
Then give 5 concrete revision recommendations, prioritized by impact.
Text: …

Critical Comparison

  • Which “implicit premises” do you accept – which need to be evidenced or reformulated?
  • Where are causal claims actually only temporal sequence?

Result - Argument map + 2 targeted revisions in the text (with marking of what logically improved)

7. Proofreading + Translator

Goal

Improve language and register without changing argumentative substance; optionally transfer terminology consistently into a second language.

Task A: Proofreading (minimally invasive)

Role: Proofreading.
Improve grammar, style, and precision, but change no content statements.
Work minimally invasive:
- Replace vague words (e.g., "important," "shows") only if unambiguous.
- Preserve technical terms; if uncertain: mark [TERM?].
At the end, give a list of the 5 most frequent pattern errors you corrected.
Text: …

Task B: Translating (optional)

Role: Translator.
Translate the text into English (or French/Spanish).
Conditions:
- Preserve register (academic) and hedge markers (e.g., "could," "suggests").
- Keep proper names and quotes unchanged.
- Give a mini-glossary table (5–10 term pairs) for key terms.
Text: …
Target language: …

Result - Corrected version + (optional) translation + glossary

8. Abstract Writer

Goal

Distill the text to its supporting statement: research problem, approach, finding/argument, relevance.

Task

Role: Abstract writer.
Write an abstract of 120–160 words.
Must contain:
- Research problem (1 sentence)
- Approach/material (1 sentence)
- Core argument/finding (2–3 sentences)
- Contribution/relevance ("so what?") (1 sentence)
No new facts; if unclear: mark [UNCLEAR].
Text base: …

Quality Check

  • Does the abstract cover the actual argumentative point or only the topic area?
  • Does the claimed method match what you actually do (source work vs. literature discussion)?

Result - Abstract + 2-sentence meta-comment: What did AI wrongly emphasize/omit?

9. Comprehension Checker (Re-Run)

Goal

Test readability and interpretation margins from the target audience’s perspective.

Task

Role: Comprehension checker.
Read my text as a critical reader in a seminar.
1) What is the thesis in one sentence?
2) Which 3 points are unclear or ambiguous (each with quote location)?
3) Which passage seems unsupported?
4) Which two questions would you ask me in feedback?
Text: …

Result - 3 refined passages in the text (1–2 sentences revised each)

Learning Outcome

At the end, you have a small package of artifacts that is scientifically reusable:

  • Outline (with evidence gaps)
  • 1–3 revised paragraphs (with traceable revisions)
  • Argument map (logic check)
  • Abstract (and optionally translation + glossary)
  • AI protocol (prompt and decision documentation)

Further Resources

Bibliography

Bender, Emily M., Timnit Gebru, Angelina McMillan-Major, and Shmargaret Shmitchell. 2021. “On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? 🦜.” In Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency (FAccT ’21), 610–23. New York, NY, USA: Association for Computing Machinery. https://doi.org/10.1145/3442188.3445922.
Campbell, Chris. 2025. “The Historian in the Age of AI.” Transactions of the Royal Historical Society, December. https://doi.org/10.1017/S0080440125100509.
Oberbichler, Sarah, and Cindarella Petz. 2025. “Working Paper: Implementing Generative AI in the Historical Studies,” February. https://doi.org/10.5281/zenodo.14924737.
The Turing Way Community. 2022. “The Turing Way: A Handbook for Reproducible, Ethical and Collaborative Research.” Zenodo. 2022. https://doi.org/10.5281/zenodo.3233853.
WarningAutomated Translation Disclaimer

This exercise was automatically translated from German using AI and may contain errors or inaccuracies. Please refer to the original German version for the authoritative text. If you notice any translation issues, please report them.

Back to top

Citation

BibTeX citation:
@inreference{mähr2025,
  author = {Mähr, Moritz},
  title = {Academic {Writing}},
  booktitle = {Critical AI Literacy for Historians},
  date = {2025-12-24},
  url = {https://maehr.github.io/critical-ai-literacy-for-historians/en/exercises/writing.html},
  langid = {en}
}
For attribution, please cite this work as:
Mähr, Moritz. 2025. “Academic Writing.” In Critical AI Literacy for Historians. https://maehr.github.io/critical-ai-literacy-for-historians/en/exercises/writing.html.