A research workflow with prompts

AI Research Project Workflow

Ten steps from defining the problem to writing up the results, with a focus on avoiding data leakage, reproducibility and explanation.

  • Steps: 10
  • Difficulty level: Expert
  • Estimated time: 3 to 12 months.
  • Target audience: Master's and PhD students and researchers in machine learning and its applications.

A workflow for an experimental project in machine learning or deep learning. It uses coding agents to build and review the project, and puts methodological safeguards at every stage: splitting, baselines, ablation and explanation.

You can copy the prompt of any step and use it on its own without completing the workflow. The “I finished this step” mark is saved in your browser only.

  1. Step 1: Defining the problem

    3 days

    State the applied research problem clearly.

    Required inputs

    • The idea and the gap

    Expected output

    • A research problem

    Step prompt

    Writing a Research Problem Statement

    Builds a coherent research problem from five elements: context, the ideal situation, the reality, the gap and the impact, flagging the places that need sources.

    Preparing the plan and proposalIntermediate
    Show the prompt text here
    You are an academic supervisor in [Specific field] who reviews the research proposals of [Stage or type of work] students.
    
    Help me formulate the research problem for this topic:
    [Research topic]
    
    The research gap I identified:
    [Research gap]
    
    [Documented evidence or figures you have]
    
    Required:
    1. Formulate the research problem using the structure: (a) general context, (b) desired situation, (c) current reality and what is missing, (d) the specific gap, (e) the impact of not addressing it, (f) what this research will contribute.
    2. Give two versions: a short one (120 to 150 words) and a detailed one (300 to 400 words).
    3. Put [source needed] after every factual or numerical claim for which I have not provided evidence.
    4. End with a list of 5 critical questions a defence committee member might ask about this wording.
    
    Academic integrity rules (mandatory):
    - Do not invent references, author names, DOIs, links, experimental results or statistics. Where a source is needed, write [source needed].
    - Do not claim to have read a study or file whose text I have not given you.
    - Clearly separate what comes from the material I provided from your own conclusions and suggestions.
    - If you are missing essential information, ask me for it before continuing instead of assuming it.
    
    Write the answer in [Answer language].
    Customise the variables on the prompt page

    Research tools for this step

    Tools linked to this step's prompt.

    Next step
  2. Step 2: Related work and baselines

    2 weeks

    Compare the existing methods to choose the baselines.

    Required inputs

    • Papers on the task

    Expected output

    • A methods matrix
    • A list of baselines

    Step prompt

    Show the prompt text here
    You are an expert researcher in [Specific field], experienced in the critical synthesis of literature.
    
    My topic: [Research topic]
    
    The studies (numbered abstracts or reading cards):
    <studies>
    [Studies]
    </studies>
    
    Required:
    1. A literature matrix as a table with the columns: [Matrix columns].
    - Write "not reported" when the information is missing, and do not guess it.
    2. Analysis of the matrix:
    - Patterns of agreement (with the study numbers).
    - Contradictions and their possible explanations (differences in data, method, metrics).
    - Trends over time in methods or results.
    - What all the studies overlook.
    3. Propose a thematic structure (not study by study) for writing the review, based on the matrix.
    
    Academic integrity rules (mandatory):
    - Do not invent references, author names, DOIs, links, experimental results or statistics. Where a source is needed, write [source needed].
    - Do not claim to have read a study or file whose text I have not given you.
    - Clearly separate what comes from the material I provided from your own conclusions and suggestions.
    - If you are missing essential information, ask me for it before continuing instead of assuming it.
    
    Write the answer in [Answer language].
    Customise the variables on the prompt page

    Research tools for this step

    Next step
  3. Step 3: Exploratory data analysis

    1 week

    Examine the data and spot possible leakage before modelling.

    Required inputs

    • Data description

    Expected output

    • An EDA report
    • Code

    Step prompt

    Show the prompt text here
    You are a senior data scientist in [Specific field].
    
    Description of the data:
    - Source and purpose: [Data source and purpose]
    - Approximate number of rows: [Approximate number of rows]
    - Columns, their types and meanings:
    [Columns and their types]
    - Target variable (if any): [Target variable]
    - A sample of 5 rows (after removing any personal data):
    [Sample rows]
    
    Required:
    1. An ordered exploratory analysis plan: quality (missing values, duplicates, impossible values), distributions, relationships, the target variable and class imbalance, and time analysis if there is a time column.
    2. Checks for possible data leakage specific to this kind of data (for example, several rows for the same patient, columns calculated after the target occurs).
    3. Complete Python code (pandas and matplotlib or seaborn) that runs in Colab, with comments, reading the file from a path I specify: [File path]
    4. A list of the questions I should answer after running the code, and what each possible output means.
    
    Academic integrity rules (mandatory):
    - Do not invent references, author names, DOIs, links, experimental results or statistics. Where a source is needed, write [source needed].
    - Do not claim to have read a study or file whose text I have not given you.
    - Clearly separate what comes from the material I provided from your own conclusions and suggestions.
    - If you are missing essential information, ask me for it before continuing instead of assuming it.
    - Do not assume the results of the analysis; the code runs on my actual data.
    
    Write the explanation in [Answer language], and the code and its comments in English.
    Customise the variables on the prompt page

    Research tools for this step

    Next step
  4. Step 4: Experimental protocol

    1 week

    Design the splitting, metrics, repetitions and statistical tests.

    Required inputs

    • The task and the data

    Expected output

    • An experimental protocol

    Step prompt

    Show the prompt text here
    You are a senior machine learning researcher and a reviewer for conferences in the field, specialising in [Specific field].
    
    Task: [Task]
    Data: [Data description]
    Proposed models: [Proposed models]
    Computing resources: [Computing resources]
    
    Design the experimental protocol:
    1. Data splitting: the method (at patient, user, site or time level) and why, the proportions, and an external test set if possible.
    2. Data leakage check: a specific list for this data (duplicate samples, preprocessing before splitting, selection on the test set, future information).
    3. Baselines: simple, strong and recent, ensuring a fair comparison (the same tuning budget).
    4. Metrics: primary and secondary metrics suited to class imbalance, and calibration metrics if the application is medical.
    5. Hyperparameter tuning: the method, search space and budget, and on which set.
    6. Repetition and uncertainty: number of seeds, confidence intervals (bootstrap), and the statistical test for comparing models.
    7. Error analysis and subgroup analysis (fairness).
    8. A reproducibility checklist: versions, seeds, settings and hardware.
    9. The empty results table that I will fill in.
    
    Academic integrity rules (mandatory):
    - Do not invent references, author names, DOIs, links, experimental results or statistics. Where a source is needed, write [source needed].
    - Do not claim to have read a study or file whose text I have not given you.
    - Clearly separate what comes from the material I provided from your own conclusions and suggestions.
    - If you are missing essential information, ask me for it before continuing instead of assuming it.
    - Do not predict performance figures.
    
    Write the answer in [Answer language].
    Customise the variables on the prompt page

    Research tools for this step

    Tools linked to this step's prompt.

    Next step
  5. Step 5: Ablation study plan

    2 days

    Decide in advance which components you will test.

    Required inputs

    • The model architecture

    Expected output

    • An ablation matrix

    Step prompt

    Show the prompt text here
    You are a deep learning researcher in [Specific field].
    
    The proposed model and its components:
    [The model and its components]
    
    Current result of the full model: [Result of the full model]
    Computing budget: [Computing budget]
    
    [Ablation results]
    
    Required:
    1. A list of the components that can be ablated, and the hypothesis that ablating each one tests.
    2. An experiment matrix: removing one component, replacing it with a simple alternative, and only the necessary interaction experiments; ordered by priority within the budget.
    3. Fairness controls: the same settings and seeds, and the same number of parameters or compensation for it.
    4. If I provide results: interpret them taking uncertainty into account, and identify any component whose effect is not demonstrated.
    5. The ablation table in the usual format used in papers.
    
    Academic integrity rules (mandatory):
    - Do not invent references, author names, DOIs, links, experimental results or statistics. Where a source is needed, write [source needed].
    - Do not claim to have read a study or file whose text I have not given you.
    - Clearly separate what comes from the material I provided from your own conclusions and suggestions.
    - If you are missing essential information, ask me for it before continuing instead of assuming it.
    
    Write the answer in [Answer language].
    Customise the variables on the prompt page

    Research tools for this step

    Tools linked to this step's prompt.

    Next step
  6. Step 6: Building the project with a coding agent

    1 week

    Build an organised, reproducible project.

    Required inputs

    • The protocol

    Expected output

    • An organised repository
    • README

    Step prompt

    Show the prompt text here
    You are a research software engineer working as a coding agent inside my repository.
    
    Project: [Project description]
    Language and framework: [Language and framework]
    The operating system I work on: [Operating system]
    [Existing code]
    
    Work in stages, and stop after each stage to show what you did:
    1. Propose the folder structure (src, configs, scripts, tests, notebooks, data with a README explaining the data source without uploading it) and wait for my approval.
    2. Set up environment management (requirements.txt or pyproject with pinned versions) and installation instructions for [Operating system].
    3. A configuration system (YAML or Hydra) so that each experiment is defined by a configuration file.
    4. A training and evaluation script that logs: the configuration, the seed, the code version (git hash) and the metrics to a file.
    5. Fix the random seeds and split the data in a fixed, saved way.
    6. Basic unit tests for data loading and the metrics.
    7. A README with the exact commands to reproduce each experiment.
    
    Strict constraints:
    - Actually run every command or test before reporting that it succeeded, and show the output. If you cannot run it, say so explicitly.
    - Do not delete existing files and do not modify the raw data.
    - Do not put API keys or passwords in the code.
    - Ask me about any unclear architectural decision.
    
    Write your explanation in [Answer language].
    Customise the variables on the prompt page

    Research tools for this step

    Tools linked to this step's prompt.

    Next step
  7. Step 7: Code review

    2 days

    Check the code for leakage and silent bugs before the final experiments.

    Required inputs

    • The repository

    Expected output

    • A list of critical fixes

    Step prompt

    Show the prompt text here
    You are an expert code reviewer for research software and machine learning.
    
    Review the following code (or the repository files if you are an agent working inside it):
    <code>
    [Code]
    </code>
    
    Experiment context: [Experiment context]
    
    Review in the following order, and for each comment give: the file and line, the problem, why it affects the results, and the suggested fix:
    1. Data leakage: preprocessing, normalisation or feature selection before splitting, overlap between sets, using the test set for tuning or early stopping.
    2. Correctness of the metrics: averaging (macro/micro), thresholds, computing metrics per batch instead of over the whole set.
    3. Randomness and reproducibility: seeds, non-deterministic operations, data ordering.
    4. Silent bugs: array shapes, model.eval() or no_grad, class order, different normalisation between training and testing.
    5. Then performance and style, briefly.
    
    Classify each comment: critical (changes the results), important, improvement.
    
    Constraints:
    - Do not mention a problem whose location in the code you cannot point to.
    - If you are an agent and you modified the code, run the tests and show their output, and do not claim a success you have not verified.
    
    Write the answer in [Answer language].
    Customise the variables on the prompt page
    Next step
  8. Step 8: Evaluation report

    1 week

    Turn the actual results into an honest results section.

    Required inputs

    • Experimental results

    Expected output

    • Results tables and paragraphs

    Step prompt

    Writing a Model Evaluation Report

    Turns your actual experimental results into an organised results section: comparison tables, statistical significance, error analysis and limitations, without embellishment.

    Analysis and experimentsAdvanced
    Show the prompt text here
    You are a senior machine learning researcher writing the results section of a paper in [Specific field].
    
    Brief experimental setup: [Experimental setup]
    
    The actual results (copied tables or logs):
    <results>
    [Actual results]
    </results>
    
    Error analysis notes (if any): [Error analysis notes]
    
    Required:
    1. A formatted main comparison table (mean ± standard deviation or confidence interval as reported), highlighting the best only when the difference is outside the uncertainty range.
    2. An analytical paragraph for each main result: the actual size of the difference and its practical meaning.
    3. Error and subgroup analysis based on my notes.
    4. Negative or mixed results, stated clearly.
    5. Suggested figures (ROC or PR curves, confusion matrix, calibration) and the message of each.
    6. A paragraph on the limitations of the evaluation.
    
    Academic integrity rules (mandatory):
    - Do not invent references, author names, DOIs, links, experimental results or statistics. Where a source is needed, write [source needed].
    - Do not claim to have read a study or file whose text I have not given you.
    - Clearly separate what comes from the material I provided from your own conclusions and suggestions.
    - If you are missing essential information, ask me for it before continuing instead of assuming it.
    - Do not calculate new figures, and do not round figures in a way that changes the ranking.
    
    Write the answer in [Answer language].
    Customise the variables on the prompt page

    Research tools for this step

    Tools linked to this step's prompt.

    Next step
  9. Step 9: Explanation (XAI)

    1 week

    Add explanations evaluated for faithfulness.

    Required inputs

    • The trained model

    Expected output

    • Evaluated explanations

    Step prompt

    Selecting and Evaluating XAI Methods

    Chooses explanation methods suited to your model and data type (Grad-CAM, SHAP, IG...), designs an evaluation of the explanations' faithfulness, and warns against misleading explanations.

    Analysis and experimentsExpert
    Show the prompt text here
    You are a researcher in explainable AI (XAI) in [Specific field].
    
    Model: [Model]
    Data type: [Data type]
    Target audience for the explanation: [Explanation audience]
    Explanation question: [Explanation question]
    
    Required:
    1. Identify the type of explanation needed (local or global, post hoc or intrinsic) based on the question and the audience.
    2. Suggest 2 to 3 suitable methods, and for each: how it works, its requirements, its known weaknesses, and the common library for applying it.
    3. A plan for evaluating the explanations: faithfulness checks (such as parameter randomisation tests, deletion and insertion), stability under perturbation, and human evaluation with experts if possible.
    4. How to present the explanations in the paper without a misleading impression (representative, not cherry-picked, samples; failure cases).
    5. Short Python code for one suggested method, with comments.
    
    Academic integrity rules (mandatory):
    - Do not invent references, author names, DOIs, links, experimental results or statistics. Where a source is needed, write [source needed].
    - Do not claim to have read a study or file whose text I have not given you.
    - Clearly separate what comes from the material I provided from your own conclusions and suggestions.
    - If you are missing essential information, ask me for it before continuing instead of assuming it.
    - Name only well-known methods and libraries, and remind me to check their APIs in their current versions.
    
    Write the explanation in [Answer language], and the code in English.
    Customise the variables on the prompt page

    Research tools for this step

    Tools linked to this step's prompt.

    Next step
  10. Step 10: Writing the paper or chapter

    2 to 4 weeks

    Write up the methodology and results from the outputs of the previous steps.

    Required inputs

    • The protocol and the results

    Expected output

    • Draft paper or chapter

    Step prompt

    Show the prompt text here
    You are an expert academic editor in [Specific field].
    
    Write a draft of the "[Required section]" section of a [Document type].
    
    The raw material (my notes and results, the only source of content):
    <notes>
    [Your notes and results]
    </notes>
    
    The required structure of the section (if any):
    [Section structure]
    
    Writing rules:
    1. Do not add any information, figure or result that is not in <notes>.
    2. Where a reference is needed write [source needed], and where information is missing write [missing information: ...].
    3. Keep to academic style: paragraphs that start with a topic sentence, clear transitions, and verb tenses suited to the section.
    4. Approximate length: [Approximate length] words.
    5. After the draft: a list of the marked places, and three suggestions to strengthen the section.
    
    Academic integrity rules (mandatory):
    - Do not invent references, author names, DOIs, links, experimental results or statistics. Where a source is needed, write [source needed].
    - Do not claim to have read a study or file whose text I have not given you.
    - Clearly separate what comes from the material I provided from your own conclusions and suggestions.
    - If you are missing essential information, ask me for it before continuing instead of assuming it.
    
    Write the answer in [Answer language].
    Customise the variables on the prompt page

    Research tools for this step

    Tools linked to this step's prompt.