Uploading a paper to an AI system and asking it to “summarize this article” can produce a polished overview within seconds. But a summary is not the same as understanding. Does the paper actually support your hypothesis? Can its methods and data be compared with ten other studies using the same fields? Can every number and conclusion be traced back to the original text?
If the answer is no, AI has only shortened the paper; it has not yet become part of the research workflow. Effective use of AI in science depends on three tests: Is the task clearly defined? Can the output be used in the next research step? Can the result be independently verified?
This guide follows the complete workflow from literature search and structured reading to research-gap analysis, experimental design, data analysis, manuscript preparation and pre-submission review. Each stage includes an actionable prompt and an acceptance criterion.
A broad request such as “find papers on photothermal synergistic catalysis” may generate useful clues, but it rarely produces a systematic search strategy. A better first step is to decompose the topic into searchable dimensions: research object, reaction system, energy field, catalyst material, critical variables and mechanism.
Prompt template
I am preparing a study on photothermal synergistic catalysis. Do not recommend papers or generate citations yet. First, decompose the topic into searchable research dimensions, including research object, reaction system, energy field, catalytic material, critical variables and possible mechanisms. For each dimension, provide Chinese and English keywords, common synonyms and Boolean search combinations, and explain what each combination is intended to retrieve. Mark uncertain terms as “to be verified.”
AI is useful for expanding keywords, identifying synonyms and constructing search strings. It cannot replace Web of Science, Scopus, Google Scholar, PubMed, CNKI or another authoritative database, and it cannot guarantee that a generated citation exists. Use AI to design the strategy, perform the search in a real database and verify important details in the original paper.
Acceptance criterion: The output reveals previously overlooked keywords, dimensions or combinations, and the strategy can be reproduced in a real literature database.
Summarizing one paper at a time can improve reading speed, but inconsistent summary structures leave the papers disconnected. A research map emerges only when multiple papers are extracted into the same schema.
Prompt template
I need to compare a set of papers on photothermal synergistic catalysis. Use only the papers I provide. Extract the following fields consistently: research question, catalytic system, material, experimental conditions, critical variables, evaluation metrics, main results, mechanistic explanation, limitations and future directions stated by the authors. Mark information that is absent from the paper as “not reported” and do not infer it. Whenever possible, attach the relevant page, figure, table or section to each key conclusion.
Acceptance criterion: All papers fit the same field structure, important claims are traceable to the source, and missing information is recorded honestly. A complete-looking table is unsafe if the model has filled gaps with unsupported content.
Once a structured map exists, AI can help locate discrepancies, conflicting conclusions and weak evidence chains. Avoid asking only, “What innovative topic should I study?” A topic that appears unstudied may have limited value, poor feasibility, high cost or unpublished negative results. The first question should be where the current evidence is incomplete.
Prompt template
Based on the literature table I provided, do not propose innovation directions immediately. Identify four types of difference: research combinations not yet covered, variables not compared systematically, conflicting conclusions and weak links in existing evidence chains. For each candidate question, list the available evidence, missing evidence, relevant papers, possible reasons the topic has remained understudied, potential value and experimental feasibility risks. Mark insufficient evidence explicitly; do not fill gaps with speculation.
Acceptance criterion: Each candidate question can be traced to specific papers and evidence, with a clear relationship between what is known and what remains missing.
A request such as “design a photocatalysis experiment” often produces a plausible but generic protocol that ignores available equipment, material constraints, measurement accuracy, safety and cost. Start from the hypothesis, define the evidence required to test it, and then map each evidence item to an experiment.
Prompt template
My research hypothesis is [hypothesis]. Available equipment: [equipment]. Available materials: [materials]. Constraints: [time, accuracy, safety or cost]. Target metrics: [metrics]. Do not generate a complete protocol yet. First list the evidence required to test the hypothesis. Then design an experiment for each evidence item and check the blank controls, comparison groups, replicates, confounding variables and measurement error. Finally, identify where the design is most likely to produce an incorrect conclusion and which parameters require confirmation by a domain expert.
AI can identify missing variables and controls, but it cannot observe every condition in the laboratory or replace qualified review of safety, ethics and feasibility. Chemical safety, clinical research, animal studies, sensitive samples and regulated protocols must be reviewed by authorized personnel and comply with institutional requirements.
Acceptance criterion: Every important hypothesis has a corresponding evidence requirement, every evidence item has a corresponding experiment, and major confounders are explicitly identified.
AI can organize datasets, generate analysis code, perform basic statistics, create plots and flag anomalies. The main risk is collapsing observation, statistical association and causation into one statement. “Yield increased as temperature increased” is an observation. Whether the relationship is significant requires statistical analysis. Whether temperature caused the increase requires experimental design and an evidence chain.
Prompt template
Here are the experimental design, data dictionary, raw data and analysis objective. Divide the output into four layers: facts directly observable in the raw data; statistical results; possible explanations based on those results; and causal conclusions that the current evidence cannot support. For every result, state the data columns, processing steps, statistical method and uncertainty. Flag anomalies without deleting them, and identify the original record that requires manual review.
If AI generates code, retain the code version, input data, parameter settings and output files so the complete analysis remains reviewable and reproducible.
Acceptance criterion: Every number is traceable to a raw record, every figure to a processing step, every statistical claim to a method and parameter set, and every scientific interpretation is clearly labeled as observation, inference or evidence-supported conclusion.
AI can help structure a manuscript, compress an abstract, edit language, organize captions and detect repetition. However, asking it to “write a paper from these results” invites the model to bridge missing material with fluent but unsupported prose. Build a claim–evidence matrix before drafting or polishing.
Prompt template
The following materials contain the research question, methods, raw results, statistical results and verified references. Do not polish the manuscript yet. First extract every core claim the manuscript intends to make and match it to the supporting data, figure, experiment and citation. Mark claims without direct support as “insufficient evidence,” and list possible overstatements separately. After completing the evidence check, improve the structure and language without adding facts, inventing data or generating new citations.
Experiments that were not performed cannot be described as results. Missing data cannot be supplied by AI. References that have not been checked against the original source should not enter the bibliography. Unpublished manuscripts, patient information, commercial data and other restricted material should not be uploaded to unapproved services.
Acceptance criterion: Every fact, number, citation and conclusion is traceable, and the model maintains an explicit boundary wherever the evidence is incomplete. AI can assist with writing; authorship responsibility cannot be outsourced.
A generic request such as “find problems in this manuscript” often produces stock advice. A useful simulated review should connect every challenge to a particular claim, paragraph or experiment and distinguish mandatory corrections from optional improvements.
Prompt template
Review this manuscript as a submission candidate. Challenge it from four perspectives: novelty, experimental design, mechanistic evidence and logical completeness. For every issue, identify the relevant paragraph or core claim, current evidence, missing evidence, severity and whether it should be addressed by an additional experiment, reanalysis or narrower wording. Separate mandatory issues from optional suggestions. Do not fabricate defects or citations to increase the number of comments.
Acceptance criterion: The review prompts the researcher to recheck data, identify overstatement, add necessary explanation or narrow conclusions whose support is insufficient.
Task Definition → Research Context → Structured Output → Validation
Reusable master prompt
The research task I need to complete is [task]. I will provide the research objective, available material, constraints and current questions. First restate your understanding of the task and identify information gaps. Then produce the output using [required fields or structure]. Separate facts, inferences, recommendations and uncertainty. Attach an evidence source or validation method to every key conclusion. Do not fill in information that is absent from the material. Mark uncertain points explicitly and list the items that require human verification.
The closer a task is to information processing, the more AI can help. The closer it is to scientific judgment, the more important human responsibility becomes. AI is well suited to keyword expansion, information organization, literature comparison, structural review and first-pass analysis. Researchers remain responsible for scientific value, experimental authenticity, causal interpretation, ethical and safety compliance, and final conclusions.
AI may not replace researchers. Its deeper value is that it can expand the space researchers are able to explore—while disciplined task definition, evidence control and verification keep that exploration scientifically reliable.
Related reading: Traceable High-Throughput Photochemistry: Screening to Pilot Scale · How to Choose a High-Throughput Photochemical Reactor
Sales-Email: network@perfectlight.cn
Recommended
news
