Download PDF

AI tools cheat sheet

How to use AI in your research · COACT 2026 · Dr Luke Philbrooks · l.a.philbrooks@qmul.ac.uk

1 of 2 · Tools

01 Chatbots

  • ClaudeMy main one: writing, analysis, long documents. Projects keep files and instructions together.
  • ChatGPTAll-rounder with voice, images and a large app and connector directory.
  • GeminiSearch grounding and Google Workspace. Gems are its saved instructions.
  • Google AI StudioVery large documents and new models. Free tier trains on your data.
  • Microsoft CopilotQMUL's official tool. Use it for anything sensitive.

02 Finding and reading papers

  • ElicitSearch and pull methods and sample sizes into a table. Check a few rows.
  • Consensus"What does the evidence say about X?" Imports your Zotero library.
  • PerplexityQuick searches and deep research. Open the sources: they don't always say what it claims.
  • Research RabbitAdd one paper, get a map of related work.
  • SciteShows whether later papers support or contradict a paper.

03 Knowledge and notes

  • Gemini Notebook was NotebookLMQuestions about papers and manuals you upload; points to where it found each answer. Only knows what you give it.
  • ZoteroReference manager. Zotero MCP lets an AI search your library. It can write too: back up first.
  • ObsidianMarkdown notes. mcp-obsidian connects them to an AI.

04 Agents and coding

Claude Code · Codex · Gemini CLI Agents that work inside a project folder: read files, run code, edit. Permissions are yours to set.
Antigravity · Cursor · VS Code Editors with agents built in. More setup, easier to miss edits.
MCPconnectors Open standard for plugging AI into other software. Official ones first; check who made the rest.
Agent Skillsreusable know-how Corrected it twice? Make it a skill. Read any skill's code before installing.

Find skills: skills.sh · anthropics/skills · K-Dense scientific skills  |  Awesome lists: MCP servers · main list

05 Local models

  • OllamaCommand-line, one-step hosting of open models.
  • LM StudioPoint-and-click app for finding and running them.
  • Hugging FaceWhere open models and datasets live.

Nothing leaves your machine, but a laptop only runs small, weaker models.

06 Landmark papers

Word2vec: embeddings
Attention Is All You Need
Chain-of-thought prompting
InstructGPT: training from human feedback
Direct Preference Optimisation
Found in the Middle: long context

Habits that pay off

  • CRCF: Context, Request, Constraints, Format
  • Interview me: ask it for its questions before it starts
  • Show, don't tell: two or three examples beat ten rules
  • Say what not to do: list the exclusions
  • Fresh chats: long chats degrade, so restart and hand over a note
  • Links, then open them: a real paper can still not say that
  • Check memory: read what it has saved about you and delete what's stale
  • Markdown beats PDF: convert before you feed it in
  • One task per chat: break big jobs into steps
  • Plan, check, do: make agents show a plan before they act
  • Least access: read-only and one folder, with backups
  • You're the boss: it's the intern, and you sign the claims

AI in your research: cheat sheet

COACT 2026 · Dr Luke Philbrooks · l.a.philbrooks@qmul.ac.uk

2 of 2 · Method

Three questions, every time

  1. Could I explain and defend this without the AI?
  2. Is this something my PhD is meant to teach me? Then use AI to learn it while you do it.
  3. Would I notice if it were wrong?

Offload the mechanics. Never the judgement.

Where each task sits

  • Automate: formatting, references, boilerplate code, file conversion
  • Assist: debugging, editing your draft, summarising papers you've read
  • Augment: critiquing your design, tutoring in a new field
  • Own: your question, design, interpretation and every claim you sign

A good prompt (CRCF)

  • Context: who you are, what it needs to know
  • Request: exactly what you want
  • Constraints: sources, dates, length, what to avoid
  • Format: table, headings, bullets, code only

Add: "Before you start, ask me any questions you need answered."

By research task

TaskWhere AI helpsThe trapTools
ReadingFinding and triaging papers; mapping a field; questioning papers you've readSummaries replace reading; invented or mis-attributed citationsElicit, Consensus, Perplexity, Research Rabbit
Crossing fieldsA patient tutor; loading manuals; translating jargonFluent answers in a field you can't yet judge. Check with a person in itLearning modes, Gemini Notebook
DesigningCritic: confounders, missing controls, pre-mortemLetting it design; confidently wrong maths; not understanding your statsClaude, ChatGPT, Gemini (reasoning mode)
AnalysingWriting and running code; agents with checkpointsSilent data changes; debugging loops; results you can't explainClaude Code, Codex, Antigravity
WritingDrafts from well-structured notes; structure; "reviewer 2" feedbackLosing your voice; AI-written discussion; sending unreviewed outputClaude, skills, style guides

Prompt: teach me, don't tell me

I'm new to [topic]. Teach me [concept]. Don't give me the full answer. Ask me one question at a time, check my answer, and explain only when I'm stuck. Finish by asking me to explain it back.

Prompt: critic for your protocol

Here is my protocol. Don't rewrite it. List the three most serious threats to validity, most serious first. For each, say what a sceptical reviewer would ask. Be specific; skip generic advice.

Prompt: your discussion

My results: [...]. My interpretation: [...]. What are the strongest objections? Which alternative explanations have I not addressed? Don't rewrite my text.

Give your projects a memory

  • AGENTS.md / CLAUDE.md: what it is, where the data is, conventions, what not to touch. Keep it short and current
  • Projects (Claude, ChatGPT), Gems (Gemini)
  • Skills: anything you've corrected twice
  • Connectors (MCP): only what you need

Prompt: hand-off note

Write a hand-off note for a fresh chat: project goal in two sentences; what we did and which files changed; decisions and why; open problems; next steps. Under 300 words.

Data and privacy

  • Public papers, own drafts: most tools, with care
  • Unpublished results, IP: QMUL Copilot, or a paid tool with training switched off
  • Participant data: approved systems only; check your ethics approval
  • Local models keep data on your machine

Integrity

  • Disclose AI use as your journal, funder and thesis regulations require
  • You're responsible for all of it (UKRI puts this on the researcher)
  • Keep a short log of how you used AI on each piece of work
  • Never send anyone unreviewed AI output
  • Agents: least access, backups and Git, approve deletions yourself

Read more