Toolbox

AI Assistant Prompt Library for Research Tasks

A practical, safety-focused library of reusable prompt patterns for literature review, experimental planning, code assistance, and data exploration — with expected outputs, validation checks, traceability practices, and a short governance checklist for when to require human review.

Welcome — how to use this toolbox

This toolbox gives researchers practical, traceable prompt patterns you can copy, adapt, and test with your preferred AI assistant. Each pattern shows the recommended prompt, the expected output shape, concrete checks to validate the AI's suggestion, and short governance rules for when a human must review or sign off. Use these patterns as disciplined starting points — don’t paste outputs directly into methods, protocols, or publications without verification.

What you'll find here

  • Prompt templates for common research tasks: literature summarization, experimental planning, code assistance, and data exploration.
  • Expected output structure so you can verify completeness.
  • Validation checks and reproducibility steps that reduce hallucination and hidden errors.
  • Traceability guidance: what metadata to record and how to version prompt/responses.
  • A short governance checklist to decide when human review is required.

General safety & traceability rules (apply to all prompts)

  1. Record: store the exact prompt, model name/version, temperature/parameters, date/time, user, and the full AI response in your project log or content store.
  2. Pin context: whenever you give a paper, dataset, or protocol excerpt, include a stable identifier (DOI, dataset ID, git commit, file path) and quote the exact passage used as input.
  3. Validate: never accept an AI-suggested experimental step, reagent, or parameter without an explicit human validation step and reproducibility checks.
  4. Version prompts: maintain a small prompt library with clear labels and an explanation of intended use, limitations, and required checks.
  5. Prefer conservative model settings when requesting factual or procedural outputs (e.g., low temperature), and require human review for high-risk outputs.

Prompt patterns

1. Literature summarization (concise, traceable)

When you need a compact, evidence-linked summary of a paper or set of papers.

Template prompt (replace placeholders):

"You are an expert research assistant. Here is the text to analyze (include DOI or exact excerpt): <PASTE TEXT OR IDENTIFIER>. Produce: (a) a 3–4 sentence plain-language summary; (b) a numbered list of the main claims/findings with short evidence quotes linking to paragraph numbers or page ranges; (c) the main methods used (concise); (d) dataset(s) or materials used (if present); (e) limitations the authors state or that are apparent; and (f) two suggested follow-up experiments or analyses. For any factual claim, include the shortest exact source quote and its location. If the text does not contain required information, say 'not specified' rather than guessing."

Expected output: structured sections (Summary; Claims with citations; Methods; Datasets; Limitations; Suggested follow-ups).

Validation checks:

  • Spot-check each claimed citation against the original text — confirm quoted snippets exist and locations match.
  • Check for hallucinated citations (claims referencing figures/tables that aren’t present).
  • Ask the model for the exact sentence numbers or paragraphs used for each claim and confirm them yourself.

2. Experimental planning (formal, safety-aware)

When drafting or reviewing an experiment plan. Use conservative language and explicitly request safety considerations.

Template prompt:

"You are an experimental design advisor for <FIELD, e.g., cell biology>. Goal: <CLEAR GOAL>. Constraints: <budget, timeline, biosafety level, equipment>. Provide three alternative experimental plans, prioritized by feasibility. For each plan provide: hypothesis, primary and secondary outcomes, sample-size estimate with assumptions and calculations, required materials/equipment with catalogue numbers if known, step-by-step protocol (clear enough for a trained technician), key controls, data to collect and formats, acceptance criteria, estimated time and cost, and a short risk & safety assessment. Mark any steps that would require institutional biosafety or ethics approval."

Expected output: three numbered plan blocks, each with the listed fields.

Validation checks:

  • Have a domain expert review controls, safety steps, and feasibility before lab work begins.
  • Recompute sample-size math independently (do not rely solely on the AI's arithmetic).
  • Confirm reagents/equipment availability and check SDS and safety data sheets for hazards.
  • If any step involves human/animal subjects, require formal IRB/IACUC review before action.

3. Code assistance and reproducible notebooks

For writing or reviewing analysis scripts, small tooling, or reproducible notebooks.

Template prompt:

"You are a data/code reviewer. Context: <brief project description and environment, e.g., Python 3.10, pandas 1.5, dataset sample size>. Task: provide minimal reproducible code that performs <task> and includes: (1) header with required package versions; (2) a runnable example on mocked or synthetic data; (3) short inline comments; (4) at least one unit test or assertion; and (5) instructions to reproduce results and known limitations. Return only the code and the reproduction steps in a clearly labeled block."

Expected output: a code block with dependencies, runnable example, and a test/assertion.

Validation checks:

  • Run the provided snippet in an isolated environment and confirm the test passes.
  • Pin exact dependency versions and verify behavior across the principal versions you support.
  • Scan code for insecure patterns (shell calls, eval, secrets in text).

4. Data exploration & quality checks

When you need a reproducible exploratory analysis plan and quality checks for a dataset.

Template prompt:

"You are a data scientist. Given this dataset description: <schema, row count, key variables>, produce a reproducible exploratory analysis plan: (a) summary statistics per variable, (b) missingness matrix and common patterns, (c) suggested visualizations, (d) anomaly detection strategy, (e) likely confounders to consider, and (f) a small code skeleton (R/Python) to run the checks and produce a report. Specify checks that must pass before downstream modeling."

Expected output: checklist of QC steps, sample code skeleton, and recommended visualizations.

Validation checks:

  • Run the suggested code on a sample of your data to confirm the outputs match the description.
  • Confirm that any identified anomalies are reproducible (not artifacts of sampling or display).
  • Preserve a snapshot of raw data used for the exploration and link it to the AI response for traceability.

Short governance checklist — when to require human review

Use this quick decision guide after you receive an AI suggestion.

  1. Does the AI propose a procedural or experimental change that could affect safety, compliance, or ethics? If yes — immediate human sign-off required.
  2. Does the AI produce a novel scientific claim, interpretation, or conclusion intended for publication or recommendation? If yes — independent expert review and replication required.
  3. Does the output include code or scripts that will run in production or analysis pipelines? If yes — code review, dependency pinning, and unit tests required before deployment.
  4. Does the output rely on external data or references not included as stable identifiers? If yes — obtain source material and verify citations before relying on the output.
  5. Is the suggested plan likely to require institutional approvals (IRB, IACUC, biosafety)? If yes — secure approvals before acting.

Short reproducibility/validation checklist (actionable)

  • Saved prompt + model parameters + full response in project log.
  • Source passage or dataset snapshot with stable identifier attached.
  • Independent re-computation of any numbers, sample-size calculations, or code outputs.
  • Peer review by at least one domain expert before operational changes.
  • Version-controlled test or notebook reproducing the key result.

How to adapt these prompts for your team

Start small: pilot a single prompt on a low-risk task, record the results, and run the validation checklist. Encourage teams to keep a short metadata record for each prompt (purpose, owner, date, approved validators). Over time, collect prompts that consistently produce high-quality, verifiable outputs and promote them to standard operating prompts in your domain-specific toolbox.

Next steps & capability opportunities

Suggested practical next steps:

  • Pilot these prompts on 3 representative tasks and collect the prompt/response pairs and validation notes.
  • Integrate a lightweight prompt/response logger so every team member records metadata automatically (prompt, model, parameters, user, date).
  • Create a short living document that maps prompts to required validation and approval steps (this toolbox can be that document).

Related implementation opportunities: consider exposing these prompt templates inside an interactive prompt-builder (for reuse and versioning), capture submissions via the platform's form storage so teams can track prompt performance over time, and package the library as a tailorable domain collection for specific disciplines.

Discuss Ask the Librarian

Discussion

Comments and conversation will live here.