# AI practice for researchers

Research briefing for the first 10 minutes of the departmental presentation. Prepared 28 October 2026. Product facts were checked against official documentation on that date; workflow recommendations below are proposed laboratory practice, not vendor performance rankings.

## The message

AI is useful for preparing, transforming and testing scientific work when the researcher defines the evidence and checks the result. The unit of success is a verifiable output: a supported claim, an inspected manuscript revision, a reproducible analysis or a tested code change.

Keep five concepts separate: **provider**, **model**, **reasoning setting**, **application/workflow**, and **subscription**. A subscription called Pro is not a universal model specification. Access to a stronger model does not establish that browsing, file parsing or execution happened.

## Three ecosystems and their tradeoffs

The following task matches are practical choices to evaluate on your own papers and data, rather than a claim that one model is inherently best at a discipline.

| Ecosystem | Relevant capabilities | Good first experiment | Main tradeoff to check |
|---|---|---|---|
| **Anthropic — Claude** | Research with sources; reusable Skills; Claude Code; agentic document work associated with Cowork | Revise a section against a supplied argument outline, then request a change log and claim audit | Available modes and local/cloud access vary. Elegant prose can conceal unsupported strengthening of a claim. |
| **OpenAI — ChatGPT, Work, Codex** | Conversational assistance, sourced research, deliverable-oriented tasks and repository work | Turn a small evidence pack into a comparison table, then reproduce one numerical result | Choosing a mode/model is separate from supplying data and authorizing tools. Tool output still needs scientific interpretation. |
| **Google — Gemini** | Deep Research, file analysis, supported Google sources and custom Gems | Compare a curated set of papers and extract a source-linked matrix | Account and administrator settings affect integration; large file collections can cause missed detail. |

Official capability references: [Claude Research](https://support.anthropic.com/en/articles/11088861-using-research-on-claude), [ChatGPT Work and Codex](https://help.openai.com/en/articles/20001275-chatgpt-work-and-codex), [Gemini Deep Research](https://support.google.com/gemini/answer/15719111?hl=en).

Do not present unsupported slogans such as “Claude is best at writing”, “Gemini never hallucinates citations” or “ChatGPT is the most accurate scientist”. A useful local evaluation uses the same input pack, time budget and rubric: citation correctness, preservation of meaning, numerical correctness, revision effort and total cost. Record the model and date.

## Chat, Work and Code

| Task | Appropriate working environment | Reviewable result |
|---|---|---|
| Explain a paper paragraph or develop an outline | Conversational chat | Explanation with quoted location and uncertainty |
| Produce a literature matrix or manuscript package | Work or equivalent agentic document workflow | Files, source links, notes on missing evidence |
| Change analysis scripts or simulation software | Codex / Claude Code or equivalent repository agent | Diff, execution evidence, tests and environment details |

OpenAI distinguishes Chat, Work and Codex by their main task orientation. They are not three independent measures of intelligence. Work and Codex share execution and permission mechanisms; local execution does not imply that all context stays on the device. [Work overview](https://learn.chatgpt.com/docs/enterprise/chatgpt-work-overview).

For Claude, avoid a screenshot-dependent explanation of separate Chat/Cowork tabs. The current official documentation describes a gradual move to a unified chat and Cowork experience; accounts may display different controls. Claude Code remains relevant to repository work. [Claude experience update](https://support.claude.com/en/articles/16761823-claude-cowork-and-chat-are-one-claude).

For a researcher, the useful distinction is **explaining a task versus executing it with files and tools**. A coding agent can edit LaTeX; a work agent can run analysis. Select the environment by access, reproducibility and the artifact you need.

## Skills and reusable laboratory procedures

A skill packages task instructions, with optional references, templates and executable scripts. Supporting material can be loaded when needed. It does not retrain the base model and does not automatically grant access to files or services. See [OpenAI skills](https://developers.openai.com/plugins/concepts/skills) and [Claude skills](https://support.claude.com/en/articles/12512176-what-are-skills).

Keep these ideas distinct:

| Mechanism | Laboratory analogy |
|---|---|
| Prompt | Today's task brief |
| Project context | The experiment folder and background material |
| Skill | A standard operating procedure with checklists and scripts |
| Connector/tool | An instrument or authorized database connection |
| Model | The reasoning and generation component |

Gemini Gems provide reusable instructions and can include knowledge files. Do not imply that a Gem is automatically interchangeable with a filesystem Agent Skill or that every platform supports the same scripts and permissions. [Gem instructions](https://support.google.com/gemini/answer/15235603?hl=en).

**Example laboratory skill: manuscript claim audit.** Inputs: a draft, verified bibliography, figure captions and evidence table. Procedure: locate each scientific claim, identify its supporting source, check quantitative agreement and flag missing support. Output: a table with claim, location, evidence, concern and proposed action. Acceptance: no invented reference, no changed result and a location for every issue. Start by validating this procedure on one known manuscript before reusing it.

## Write an acceptance contract, not a vague wish

Use six ingredients: **task, context, sources, constraints, output, checks**. Add the audience and time budget when they matter. Ask for explicit assumptions and a concise justification; requesting a long internal thought process is not a substitute for evidence.

Weak request: “Improve my paper.”

Useful request:

> Edit only the attached Introduction for clarity for a condensed-matter journal. Preserve every numerical value, citation key and scientific claim. Use the supplied terminology list. Do not add literature or strengthen causal statements. Return the revised section, a short change log, and a table of claims whose wording or support needs my decision. If a needed source is absent, mark it as missing.

Run a small first pass, inspect it, then expand. For long tasks, ask the agent to save a short record of decisions, input versions and unresolved issues. This is more reliable than repeatedly pasting an expanding conversation into a new task.

## Attach a curated evidence pack

Use a short attachment manifest that explains what each file is authoritative for. Separate evidence from desired writing style and from preliminary ideas.

| File | Purpose | Useful annotation |
|---|---|---|
| manuscript_v07.tex or .docx | Text to revise | Allowed sections; protected claims; version |
| references_verified.bib | Approved reference metadata | Which full texts are also supplied |
| paper_01.pdf … paper_05.pdf | Scientific evidence | DOI, year, relevant sections; preprint/published status |
| results.csv | Numerical evidence | Units, columns, missing-value conventions |
| figure_03.png + caption.txt | Figure interpretation | Axis units, color scale and original data path |
| task_brief.txt | Acceptance criteria | Deliverables, scope and stop conditions |

Ask the model to list the files it can actually read and identify missing pages, unreadable equations or ambiguous columns before synthesis. Uploading a document is not proof that every page or figure was interpreted. When a figure matters, provide it explicitly with its caption and, when possible, its numerical data. Use OCR for scanned pages and inspect the extracted equations. Prefer manageable thematic batches over an undifferentiated archive.

Supported inputs and limits differ across products and plans. The current [ChatGPT upload guide](https://help.openai.com/en/articles/8555545-uploading-files-and-audio-to-chatgpt) and [Gemini file guide](https://support.google.com/gemini/answer/14903178?hl=en) should be checked before a live demonstration. Avoid putting rapidly changing file limits on the main slides.

## Literature research as a traceable pipeline

1. Define the scientific question, date range, materials, methods and inclusion/exclusion criteria.
2. Generate search terms and synonyms; search an appropriate scholarly index and primary sources. Use AI research for discovery and synthesis, while recording the actual search coverage.
3. Deduplicate records by DOI and title. Check the publisher record; distinguish preprints, versions and corrections.
4. Read the relevant full-text passage. An abstract or search snippet supports only what it contains. Mark inaccessible full text explicitly.
5. Build an evidence table before writing a narrative. Include negative findings and disagreements.
6. Verify every reference and central scientific claim yourself. Export citations to a reference manager and retain the search log.

Deep Research products can gather and synthesize multiple sources, but a sourced report is not automatically an exhaustive systematic review. For ChatGPT, the research plan and source selection are reviewable. [Deep research guide](https://help.openai.com/en/articles/10500283-deep-research-in-chatgpt).

**Prompt for the literature step**

> Identify primary studies of spin-wave propagation in patterned YIG waveguides within the date interval I specify. Separate experimental and numerical work. For each candidate, provide title, authors, year, DOI, geometry, excitation method, damping assumption and the exact section supporting the extracted result. Mark abstract-only access. Do not infer missing numerical values. First propose the search and screening plan; then return an evidence matrix and a list of unresolved disagreements.

Suggested evidence-table columns: paper/DOI; access level; method; geometry; material parameters; boundary conditions; central result; limitations; supporting page/figure; checked by/date. This is a template, not a claim that a search has already been performed.

## Manuscripts with a controlled change budget

Work in separate passes: argument structure, scientific consistency, language, references and formatting. Start from author-written results and a verified evidence table. Preserve a versioned original and review a diff or tracked changes. A convincing sentence is not a new result.

**Prompt for scientific consistency**

> Audit the supplied Methods and Results without rewriting them. Check symbol definitions, SI units, parameter consistency, boundary conditions, figure references and whether conclusions exceed the reported results. Return issue, exact location, evidence and severity. Distinguish a demonstrated error from a question requiring the authors' judgement. Do not invent experiments or citations.

**Prompt for a reviewer response**

> Build a response table from the supplied reviewer comments and our revision notes. Columns: comment, proposed response, manuscript change, evidence and unresolved decision. Claim an experiment or revision was completed only if it appears in the supplied records. Keep the tone factual and respectful. Leave missing results as explicit placeholders for the authors.

Follow the target journal's current AI and confidentiality rules. Authors retain accountability; AI should not be named as an author or used to fabricate citations, data or results. Unpublished third-party manuscripts and sensitive information require an approved environment and appropriate permission. [Nature AI policy](https://www.nature.com/nature/editorial-policies/ai). Local computer access alone is not a privacy guarantee.

## When to use Pro or stronger reasoning

Distinguish **paid access**, **selected model**, **reasoning effort**, and **research/tool mode**. OpenAI's current help identifies GPT-6 Pro as powered by GPT-6 Astra; availability is plan- and workspace-dependent. This is a dated product fact, not the main teaching point. [Current model and experience guide](https://help.openai.com/en/articles/20001275-chatgpt-work-and-codex).

Use a staged workflow:

1. A fast/default pass extracts and organizes inputs.
2. A stronger reasoning pass addresses a bounded hard problem: a derivation, competing explanations, consistency across equations, or a consequential review.
3. Verify using calculation, source passages, tests or an independent reviewer.

Provide the equations, conventions and known checks before increasing reasoning effort. Ask for limiting cases, dimensional checks and alternative explanations. Do not spend premium reasoning on a missing attachment or an unspecified task.

**Prompt for a stronger reasoning pass**

> Starting from the attached energy functional and stated boundary conditions, check the proposed effective-field derivation. Use our H-field convention in A/m and define the gyromagnetic coefficient. Identify assumptions, verify SI dimensions and compare the uniform-state and zero-damping limits. Return the corrected derivation and a short list of numerical checks. Cite only the supplied sources or sources you actually retrieve.

Two models agreeing is useful triangulation, not independent experimental validation. Measure total effort, including human correction and verification, when assessing whether a stronger model saves time.

## The one-minute laboratory checklist

Before accepting an AI result: Is the source real and relevant? Are units and numerical values preserved? Was the calculation actually executed? Can the change be inspected? Is the information permitted in this environment? Is there an accountable human decision?

Transition to Fullmag: **The same principle applies to simulation software: an attractive output becomes scientific evidence only when we can trace the model, numerical method and validation.**