Act as a machine learning mentor helping me solve and understand [ML PROBLEM]. My current level is [CURRENT LEVEL], my dataset or data situation is [DATASET DETAILS], and my practical constraints are [CONSTRAINTS]. My goal is [LEARNING OR PROJECT GOAL]. First, restate the prediction or decision problem precisely: unit of prediction, target, prediction time, available-at-prediction-time features, and cost of false positives versus false negatives. Identify whether this is supervised learning, unsupervised learning, forecasting, ranking, causal inference, or another framing; explain any ambiguity rather than forcing a label. Produce:
1. A 5-8 step learning and implementation plan, with the concept I need to understand at each step and one small exercise or diagnostic.
2. A data-audit checklist covering target leakage, duplicate entities, missingness, class imbalance, time ordering, train/test contamination, and sensitive attributes.
3. A baseline-first modeling plan: simple baseline, candidate models, feature treatment, validation strategy, and metrics tied to the stated error costs.
4. Pseudocode or compact annotated code for the next step only, using [PREFERRED TOOLS], with explanations of key choices.
5. An experiment log template listing hypothesis, change, split, metric, result, and interpretation.
Do not invent dataset columns, benchmark scores, citations, or experimental results. Explicitly distinguish conceptual advice from assumptions about my data. Before answering, check that evaluation respects the real deployment timing and that no feature could leak future information. Ask up to 3 clarifying questions only if a required input is missing.Fill in
| Placeholder | What to enter | Example |
|---|---|---|
| [ML PROBLEM] | Describe the problem you want to solve, including the business or research context. | Predict whether a customer will cancel a monthly software subscription in the next 30 days |
| [CURRENT LEVEL] | State your relevant experience, such as beginner with Python or experienced analyst new to deep learning. | Comfortable with Python and pandas; new to model validation |
| [DATASET DETAILS] | Describe the rows, target, key columns, size, data source, and known quality issues. | One row per account-month, 48,000 rows, cancellation label, usage events aggregated monthly, plan type, tenure, support tickets, and renewal date |
| [CONSTRAINTS] | List limits such as runtime, interpretability, privacy, deployment environment, or deadline. | Predictions run weekly; account managers need interpretable reasons; no data collected after the scoring date |
| [LEARNING OR PROJECT GOAL] | State what you need to learn or deliver by the end. | Build a trustworthy baseline and explain validation choices to my manager |
| [PREFERRED TOOLS] | Name the languages and libraries you want to use. | Python, pandas, scikit-learn, and Jupyter |
How to use
- Describe the data-generating process, not just the file name; include when each field becomes available.
- Use the data-audit checklist before selecting a complex model or tuning hyperparameters.
- Run the baseline and record the exact split, metric, and result in the experiment log.
- Follow up with: “Here are my validation results and confusion matrix. Help me diagnose the largest errors without proposing changes unsupported by this evidence.”
Variations
Paper walkthrough
Use this to learn a research paper without mistaking a summary for replication.
Teach me this machine learning paper: [PAPER TEXT OR LINK EXCERPTS]. My background is [BACKGROUND], and I want to understand [FOCUS]. Produce a structured walkthrough of the problem, assumptions, method, objective function, experimental setup, claimed contribution, and limitations. Define notation before using it and give a small worked intuition example. Separate what the authors report from your interpretation. End with three replication questions and three ways the method could fail in a different dataset. Do not invent details absent from the paper.
Model debugging
Use this after a model performs poorly or inconsistently.
Help diagnose this ML model using [PROBLEM DEFINITION], [DATA AND SPLIT DETAILS], [METRICS], and [ERROR EXAMPLES]. Produce a ranked debugging plan covering leakage, split validity, label quality, preprocessing consistency, baseline comparison, calibration, subgroup performance, and train-serving skew. For each proposed test, state what outcome would support or weaken the hypothesis. Do not prescribe a new algorithm until simpler causes are checked. Ask up to 3 questions only when key diagnostic evidence is missing.
Interview practice
Use this to prepare for an ML system-design interview.
Run a machine learning system-design practice interview for [USE CASE] at [SENIORITY LEVEL]. Ask one question at a time, wait for my answer, then critique it against problem framing, labels, data collection, features, modeling, offline evaluation, online metrics, deployment, monitoring, and safety or fairness. After eight questions, provide a scorecard and an improved answer outline. Do not assume access to proprietary data or claim there is one correct architecture.
Tips
- Define the scoring timestamp before discussing features; this is the clearest defense against target leakage.
- Choose the validation split to mirror deployment, using time-based splits when future outcomes are predicted from past observations.
- Keep a trivial baseline such as majority class, last value, or linear model; it tells you whether complexity earns its cost.
- Inspect errors by meaningful groups and time periods, but do not infer unfairness or causality from small or unrepresentative slices.
FAQ
Can AI build my machine learning model for me?
It can explain, draft code, and help plan experiments, but you must run the work against your data and validate the results.
What metric should I use?
Choose a metric based on the decision and error costs. Accuracy is often misleading for imbalanced classes, ranking, and threshold-based actions.
Should I start with a neural network?
Usually start with a simple baseline and a validation design. A more complex model is useful only if it improves the metric that matters and remains feasible to operate.