Ten Simple Rules for AI-Assisted Research

A condensed good-practice guide for researchers using Claude. Each rule names a principle, explains the failure mode it prevents, and points to the fuller treatment.

These are not commandments. They are the distilled lessons of researchers who have used AI tools well — and of those who have not.

Rule 1: Start with context, not tasks

Before asking Claude to do anything substantive, write a CLAUDE.md — a plain-text briefing document in your project folder that tells Claude what the project is, what the data contains, and what you do not want it to decide alone. Without this, every session starts from scratch. With it, Claude works inside your project rather than answering questions about it.

A minimal CLAUDE.md takes thirty minutes. It pays back on the first session.

→ A9.markdown-project-memory, C.start-here

Rule 2: Fail quickly and cheaply

The cost of a failed first experiment is near-zero: a few minutes, no data at risk, nothing published. The cost of not starting — while you worry about whether you are doing it right — is real and accumulates. Try something small on a task you can verify in five minutes. Your questions will be sharper after one real experiment than after ten hours of reading about AI.

This is also the right strategy within a session: process three files before processing three hundred. Check the output before committing to the full run.

→ B.lifecycle.0.iteration, C.start-here

Rule 3: Scope the task before you run it

An underbounded task produces an underperforming output — and costs more time in correction than a well-scoped task costs in preparation. Before submitting a complex prompt, state the output format explicitly, the length or scope, what you do not want included, and any non-obvious constraints. "Summarise this paper" is underbounded. "Summarise the methodology section in four bullet points, citing specific page numbers, without making claims about the conclusions" is not.

The extra thirty seconds of specification saves ten minutes of iteration.

→ A.issue.bounding, C.prompts

Rule 4: Review the plan before you run consequential tasks

For any task that modifies files, produces outputs you will use in your work, or operates on data you cannot easily reconstruct — use plan mode (Shift+Tab in Claude Code) to separate the proposal from the execution. Read what Claude plans to do. Approve it. Then run it.

This is not caution for its own sake. It is the practical form of the supervision principle: consequential actions deserve a review step, even a brief one.

→ A.issue.plan-mode, A.concept.agents

Rule 5: Verify what matters; calibrate to stakes

Not every AI output needs checking. A rough first draft of a low-stakes email does not require line-by-line verification. A data extraction that will appear in a published table does. The failure mode to avoid is treating all outputs with the same verification intensity — either checking everything (exhausting, unsustainable) or checking nothing (dangerous for high-stakes work).

The practical rule: verification effort should be proportional to the consequences of error. Before relying on an output, ask: what happens if this is wrong?

→ B.trust, C.calibration, A.critical.limitations

Rule 6: Delegate execution, not interpretation

Claude is excellent at transforming, formatting, extracting, and drafting — the execution layer of research work. It is unreliable as the final judge of what something means, whether an argument holds, or what a source actually establishes. The interpretive layer — the work that requires your disciplinary expertise and your evaluative judgment — is not safely delegatable at any adoption level.

A useful test: could you defend this output in a seminar? If not, you have delegated too much.

→ A.critical.limitations, B.lifecycle.4.dataanalysis, B.ownership

Rule 7: Watch the centroid — AI pulls toward the known

Claude's training reflects the existing distribution of published work. It is fluent at the centre of any field — the established methods, the canonical sources, the consensus framings. It becomes less reliable, and more confidently wrong, at the periphery — novel questions, underrepresented traditions, recent developments past its training cutoff, and the specific gaps that make your research original.

Use AI freely for centroid work. Be more careful with frontier work. The risk is not that Claude refuses to engage with novel questions; it is that it engages fluently and incorrectly.

→ B.centroid-periphery, B.lifecycle.1.creativity

Rule 8: Build in adversarial checks

Claude will affirm the framings you bring to it. If your research question has a hidden assumption, Claude is more likely to work within it than to surface it. If your argument has a structural weakness, Claude's feedback will be gentler than a reviewer's. This is not Claude being dishonest; it is a structural property of systems trained to be helpful.

Counter-practice: explicitly ask Claude to steelman the opposing view, list every weak point, or play the hostile reviewer. Cross-check important outputs against a different model (Gemini, GPT-4). The adversarial check is the most underused practice in AI-assisted research.

→ B.epistemics, B.autonomy

Rule 9: Name the compression, protect the time

When Claude compresses a four-hour task into one hour, you have three hours back. What you do with them is your decision — and one that requires deliberate intent, because the default is that more work fills the space immediately. The tool accelerates; it does not protect your time.

Before an AI session, know what the goal is. After the session, stop. Distinguish between using Claude to compress existing work (which returns time) and using Claude to generate new work (which does not). Both are valid uses; only the first frees your schedule.

→ B.time, B.experience

Rule 10: Disclose honestly — data, credit, authorship

Three decisions face every researcher before using AI on a piece of work: Can I share this data with Claude? (Privacy and consent.) Do I need to disclose AI's role? (Funder, journal, and institutional rules.) Am I still the intellectual author? (The ownership test: can you explain, defend, and stand behind the output?)

These decisions are not bureaucratic. They are the conditions under which AI-assisted work remains genuinely yours and publishable under your name. Apply them before you begin, not after you have submitted.

→ C.ethics-compact, A.issue.privacy-security, B.ownership

A note on sequence

Rules 1–4 are setup and process rules — they determine whether the tool works at all for serious research use. Rules 5–8 are epistemic rules — they determine whether the outputs are trustworthy. Rules 9–10 are sustainability rules — they determine whether the practice is maintainable and honest over time.

You can start anywhere. But if you are getting poor results, start with Rule 1. If you are getting good results but uneasy about them, start with Rules 6 and 10.

Related