CUIN 8329 · Academic Publishing & Presenting · October 13

Dissertation Engineering

The technology half of our team presentation, showing what changes when AI and workflow design enter the writing process.

“I am not going to give you a list of ten AI tools. The more useful question is how to build a workflow where the tools reduce friction without quietly changing the project.”

Plan your dissertation journey with AI and run audits to ensure alignment.

Slide 2 · Workflow architecture

Build a workflow

Productivity improves when the handoffs between tasks become cleaner.

Map

See how the literature connects before you read.

Organize

Keep sources, notes, and claims in one place.

Think

Design a prompt spec for the task.

Draft

Writing is an iterative process.

Audit

Check claims before polishing.

Handoff

Restart clean; maintain canon.

Most writing problems are handoff problems: search→read, read→notes, notes→claims, claims→drafts, drafts→revision. The best tools solve one handoff — and know when to stop.

Discussion: where can technology reduce friction without taking over the scholarly judgment?

Slide 3 · Map

Instead of reading randomly, map the area of interest

Litmaps shows citation relationships, so you can see how connections are structured.

Start from 3–5 papers you already trust: trace citations backward and forward, find the clusters and the papers that bridge them, notice what is missing, and build a targeted reading queue.

It is not a substitute for searching or reading — and it does not decide what literature “matters” for you.

Use Litmaps for positioning, not outsourcing.

Similar tools: ResearchRabbit, Connected Papers, Semantic Scholar.

Slide 4 · Organize

Build a research archive

One place where sources, claims, and decisions stay connected.

Zotero collections and tags, annotation extraction, a synthesis matrix, a running “claims to verify” list, a revision log, a decision memo, open questions.

Every source gets a job: evidence, theory, method model, counterpoint, or framing language. AI can summarize, sort, and query, but only if traceability survives.

An AI summary is useful only if you can get back to the source.

Slide 5 · Think

Prompt specs beat “make no mistakes”

A strong prompt is a work order, not a magic phrase.

A prompt spec includes:

Artifact being used.
Task boundaries.
Target reader or venue.
Evaluation criteria.
Output format.
Multi-model verification of sources.

Example

“Using only the provided text, identify where the argument drifts, where the literature review becomes summary instead of synthesis, and where the claim exceeds the evidence. Return a revision memo with five prioritized changes.”

Narrowing the task reduces hallucination, preserves authorship, and stops the model from rewriting the project when you only needed diagnosis.

Beginners prompt for product. Advanced users prompt for decision support.

Slide 6 · Draft

Use chains, not mega-prompts

Chains make drafting iterative, one small auditable task at a time.

1Diagnose
2Reframe
3Stress test
4Revise
5Memo

Rule of thumb: improve one decision at a time, not the whole paper at once. Chains preserve control, expose errors early, and give every step a comparable output.

Slide 7 · Model selection

Pick the right model for the job

Model choice is part of productivity and resource management.

Routine → lighter model

Formatting cleanup, checklists, meeting notes, headings.

Standard → mid-level model

Outlines, paragraph flow, revision questions, comparing versions.

Complex → strongest model

Theory alignment, design tradeoffs, synthesis, reviewer strategy, high-consequence decisions.

Escalate as ambiguity, stakes, and scholarly judgment increase — and don't spend premium usage on tasks a lighter model handles.

Slide 8 · Audit

Watch for drift: productivity can quietly change the project

Topic

The paper becomes broader than the research question.

Argument

The central claim shifts without a conscious decision.

Citation

Sources added because convenient, not necessary.

Voice

Repeated polishing flattens the writer's judgment.

Method

The analysis follows the tool instead of the study.

Keep a canon

Your canon is the written statement of what the project is — current claim, evidence base, and audience. It changes only when you explicitly decide to change it. When a tool’s output disagrees with the canon, the canon wins.

Drift is not automatically bad — research evolves. The problem is invisible drift: change you never noticed or decided.

Slide 9 · Handoff

When you hit a usage limit, stop and audit

A usage limit is a built-in quality-control checkpoint.

Re-read the argument yourself.
Check the paper still answers the original question.
Verify sources and citations.
Look for unsupported claims and unintended drift.
Compare the draft against the framework and research questions.
Note what changed and why.

Then update the project map, log decisions and open questions, and start a new thread from it — long threads accumulate stale assumptions and drift.

Slide 10 · Main artifact

A markdown map of the dissertation journey

Plain text, portable, searchable, easy to version, readable by humans and AI — the project's external memory.

Example

# dissertation-map.md

## Core claim
Faculty make AI policy visible through course-level rules.

## Current design
Quantitative survey + disciplinary comparison.

## Decisions log
- 9/22: shifted from mixed methods to quantitative.
- 10/8: first committee feedback positive.

## Open questions
- How much theory belongs in the journal version?

## Next moves
- Revise manuscript frame.
- Track committee feedback.

Why it works

Keeps decisions visible
Separates project memory from polished prose
Gives AI a cleaner artifact than a whole dissertation in a prompt

Key tip: when a thread gets long or hits a limit — update the map, save it, start a fresh thread from it.

Slide 11 · Quality control

Advanced productivity includes quality control

Speed only helps if the work stays accurate, traceable, and yours.

Source grounding — “Show me the sentence this claim depends on.”
Triangulation — compare AI output against the article, not memory.
Uncertainty tagging — verified / needs checking / speculative.
No bibliography outsourcing — verify every citation yourself.
Disclosure — follow course, journal, and institutional policy.

Slide 12 · Rehearsal

Personas: use your committee as reader lenses

A persona is not an imitation of a person — it is a reader lens built from public scholarly characteristics.

Theory lens

Does the framework explain the claim? Is theory doing analytic work or just cited for show?

Methods lens

Are variables and measures aligned? Are claims proportionate to the design?

Policy / practice lens

Why does this matter beyond the site? What would practitioners do?

Slide 13 · Live activity · about 4 minutes

Spot the Drift — you run the audit

Work on your own. Decide before you tap.

Research question: How do doctoral students in a writing course describe their use of AI tools when revising drafts?

Your data: semi-structured interviews with 15 doctoral students.

Step 1 · Tap the sentence where the draft stops answering the research question

What was wrong, and why

Drift begins here: “This pattern echoes broader findings on technology adoption in higher education (cf. Venkatesh et al., 2003)…” Everything before that sentence stays inside the data — participants describing their own revising.

Drift type: citation drift, then topic drift.

Citation drift — the Venkatesh citation is adjacent and convenient, not necessary. It decorates a generalization the interviews cannot support.
Topic drift — “writing programs across the country,” “institutions that fail to develop clear AI policies,” and “renegotiating authorship” are all outside the research question. No interview asked about institutional policy; no finding supports it.

Why most readers miss it: every sentence is fluent, plausible, and correctly formatted. Drift is invisible by definition — which is why a project canon exists, not why proofreading exists.

The move that would have caught it: the canon — “My current claim is…” — answered against the stated research question, not against the draft’s momentum.

Bonus point: the drift reads better than the data section. Productivity tools produce exactly this fluency. A fluent paragraph is not a defensible claim.

Drift types absent here: voice drift and method drift — worth confirming aloud, because all five types are different failure modes.

Live results · anonymous · updates every 4s

0 responses recorded · 0 picked the onset

Where the class thinks drift begins

Drift types chosen

Slide 14 · The takeaway

The takeaway

Map the literature before drowning in it.
One research archive, one markdown project map.
Prompt specs, chained micro-steps, models matched to task complexity.
Usage limits are audit points; restart threads from the handoff.
Verify everything — preserve authorship, traceability, and disclosure.

Advanced productivity is not doing less thinking. It is making the thinking easier to manage.

Appendix · Take it with you

Prompt specs

Prompt spec — drift detection

Using only the provided text, identify where the argument has drifted from the stated
research question. Do not rewrite the section. Return:
1. the specific sentence or paragraph where drift begins,
2. why it represents drift,
3. one revision option,
4. one question I should answer before revising.

Prompt spec — source discipline

Using only the provided article and my paragraph, identify which claims are directly
supported, partially supported, or unsupported. Quote the relevant source text for each
supported claim. Do not add outside sources.

Prompt spec — committee persona

Using only the draft excerpt and the committee-member background notes, identify likely
questions this reader might raise. Do not imitate their voice, invent views, or claim
they would definitely ask these questions. Return: 3 likely concerns,
2 revision options, 1 sentence that most needs strengthening.

Prompt spec — thread handoff

Create a markdown handoff for a new thread using only this conversation. Include:
# Current project goal / # Current argument / # Decisions made / # Open questions
# Sources or claims to verify / # Drift risks / # Next 5 actions
Do not introduce new ideas.

Appendix · Sources

Sources

Each slide’s empirical claims rest on four verified published sources.

Noy, S., and W. Zhang (2023). “Experimental evidence on the productivity effects of generative artificial intelligence.” Science 381:187–192. A randomized trial of 453 professionals on writing tasks. Time fell 40 percent and quality rose 18 percent.
Walters, W. H., and E. I. Wilder (2023). “Fabrication and errors in the bibliographic citations generated by ChatGPT.” Scientific Reports 13:14045. Documents fabricated and erroneous citations in GPT-3.5 and GPT-4 literature reviews.
Dell’Acqua, F., et al. (2023). “Navigating the jagged technological frontier.” Harvard Business School Working Paper 24-013. A field experiment with 758 consultants showing large AI gains inside its capability frontier and confident errors outside it.
COPE Council (2024). “COPE position: Authorship and AI.” Committee on Publication Ethics. AI tools cannot be authors. Authors must disclose AI use and stay responsible for all content.