AI Tools for Doctoral Work
What each tool can help with, and what remains yours

A simple starting point

Choose the job before you choose the tool.

1. 01Discover

   Generate search terms, identify unfamiliar authors, and find candidate sources.

   Keep yoursSearch strategy and source selection.

2. 02Understand

   Explain a method, compare arguments, or ask questions about a bounded source set.

   Keep yoursClose reading and interpretation.

3. 03Create

   Brainstorm, outline, reorganize, or revise prose that begins with your thinking.

   Keep yoursClaims, structure, and final wording.

4. 04Check

   Test logic, surface counterarguments, and identify places that need evidence.

   Keep yoursVerification and final judgment.

The strongest use begins with a specific task and ends with a human decision.

General-purpose assistants

ChatGPT, Claude, and Gemini overlap more than they differ.

1. ChatGPTFlexible dialogue

   Useful for explanation, brainstorming, revision, and working with uploaded files.

   Try it forQuestions, critique, and transformations.

2. ClaudeDocument work

   Useful for summarizing bounded material, comparing drafts, and surfacing assumptions.

   Try it forLong-form critique and version comparison.

3. GeminiConnected formats

   Useful for questions across files and web sources, with Google connections when enabled.

   Try it forCross-format exploration and research planning.

4. Shared limitPlausible is not proven

   Any of them can omit context, invent a citation, or state an error confidently.

   Your moveVerify claims in the original source.

Capabilities and access change quickly. Choose by task, privacy, and institutional policy, not brand loyalty.

Research-focused tools

A narrower tool can be more useful than a general chatbot.

1. ElicitFind and screen

   Semantic paper search, question-specific summaries, screening, and structured extraction.

   Watch forCoverage gaps and unverified metadata.

2. NotebookLMQuestion your sources

   Answers and study materials grounded in a source set you choose, with inline citations.

   Watch forA weak or incomplete source set.

3. ZoteroKeep the record

   A non-AI source of record for collecting, organizing, annotating, and citing research.

   Use it forPDFs, metadata, notes, and references.

4. Library searchAnchor the search

   Databases, Google Scholar, citation trails, and librarians remain essential.

   Use them forCoverage, provenance, and full text.

Let AI help you find and question sources. Let the original texts and your reference manager anchor the record.

A basic coding-agent workflow

Claude Code and Codex can work with an entire project, not just one prompt.

• 1. OrientAsk it to explain the repository and locate the relevant files.
• 2. ConstrainName the goal, boundaries, permissions, and definition of done.
• 3. ActLet it make a focused edit and run appropriate checks.
• 4. ReviewInspect the diff and the live result before committing or publishing.

Public examples: proposal defense repository · live defense deck · dissertation sites hub

A coding agent can take actions. Start with a low-stakes copy, keep permissions narrow, and review the result.

Access and configuration

Pay for a recurring need. Configure for consistent work.

1. Paid accessCapacity

   Depending on the product, a plan may add higher limits, more capable models, tools, or longer agent sessions.

   Ask firstDoes the benefit justify the cost?

2. InstructionsConsistency

   Save durable preferences for voice, privacy, file boundaries, required checks, and output format.

   ExamplesCustom Instructions, AGENTS.md, CLAUDE.md.

3. Model choiceFit

   Use a faster or smaller model for routine work and a more capable model for difficult reasoning. Fastest to most capable, Claude offers Haiku, Sonnet, Opus, and Fable; ChatGPT offers Luna, Terra, and Sol.

   Rule of thumbLearn the ladder, not the version number. Start with the least costly option that works.

4. Reasoning effortDepth

   Low or light for simple tasks, medium for most work, and high or deeper for ambiguous, multi-step problems.

   RememberMore effort means more time and usage, not guaranteed truth.

Paid access should increase practical capacity, not become a proxy for rigor. Courses should still account for access and equity.

One practical workflow

Move from discovery to a decision you can explain.

• 1. SearchUse library databases and research-focused discovery tools.
• 2. SaveCapture the source, PDF, metadata, and notes in Zotero.
• 3. ReadRead the original text before relying on a generated summary.
• 4. ThinkWrite a short memo or claim in your own words first.
• 5. TestAsk an assistant for questions, counterarguments, or revision options.
• 6. VerifyCheck every kept claim, quotation, citation, and methodological detail.

AI can shorten parts of the process. It should not make the process invisible.

A reusable prompt pattern

Ask for intellectual support before asking for finished prose.

Context: I am a doctoral student examining [topic]. Here are my notes or a source excerpt.

Task: Ask me three questions that test my interpretation before suggesting changes.

Boundaries: Use only the material supplied. Do not invent citations. Flag uncertainty.

Output: Return questions, one possible counterreading, revision options, and claims I must verify.

A good prompt makes the task, evidence boundary, and verification responsibility visible.