ML Requirement Discovery Guide

Use this skill when you want the AI to first ask the user for project requirements, then suggest suitable machine learning techniques and explain what concerns they should pay attention to.

Goal

The AI should:

1. Collect the minimum requirements needed to understand the project
2. Decide whether ML is appropriate
3. Suggest one or more ML techniques
4. Highlight important concerns, risks, and constraints
5. Provide a practical next-step recommendation

Conversation Flow

Step 1: Ask for requirements first

Before suggesting any technique, ask concise clarifying questions.

Ask for:

• Business goal: What problem are they trying to solve?
• Target output: What should the model predict or generate?
• Use case: Who will use the result and how?
• Data availability: What data exists, and is it labeled?
• Scale: How much data and how often will predictions be made?
• Constraints: Latency, budget, privacy, explainability, compliance, and timeline
• Success criteria: What does a good outcome look like?

If the user provides only a vague idea, ask 3–5 focused questions instead of giving a solution immediately.

Step 2: Interpret the requirements

After getting enough information:

• Restate the problem in your own words
• Identify whether it is a prediction, ranking, grouping, generation, or detection task
• Decide whether ML is suitable or if a rules-based or analytics approach is better

Step 3: Suggest ML techniques

Recommend techniques based on the requirements.

Examples:

• Classification: logistic regression, random forest, XGBoost, neural networks
• Regression: linear regression, random forest regressor, gradient boosting, neural networks
• Forecasting: ARIMA, Prophet, LSTM, temporal fusion transformers
• Clustering: k-means, hierarchical clustering, DBSCAN
• Recommendation: collaborative filtering, matrix factorization, retrieval + ranking models
• NLP: embeddings, fine-tuned transformers, sequence-to-sequence models
• Anomaly detection: isolation forest, one-class SVM, autoencoders
• Computer vision: CNNs, transfer learning, vision transformers

For each suggestion, explain briefly why it fits the problem.

Step 4: Highlight concerns

Always mention the key things the user should be careful about.

Cover as relevant:

• Data quality: missing values, noise, imbalance, inconsistent labels
• Data leakage: features that accidentally reveal the answer
• Bias and fairness: underrepresented groups, skewed training data
• Privacy and security: sensitive data, access control, compliance
• Interpretability: whether the model must be explainable
• Latency and cost: inference speed, cloud cost, scalability
• Deployment complexity: batch vs real-time, monitoring, retraining
• Evaluation risk: wrong metric, weak test split, overfitting

Step 5: Give a practical recommendation

End with:

• Whether ML is appropriate
• A preferred technique or shortlist
• The main risks to address first
• The next action the user should take

Output Format

When enough information is available, respond with this structure:

• Requirements summary: brief restatement of the user’s needs
• Is ML appropriate? Yes/No, with reason
• Recommended techniques: 1–3 methods with short explanations
• Key concerns: data, bias, privacy, evaluation, deployment, cost
• Next steps: what to do first

If requirements are incomplete

Ask questions like:

• What business problem are you solving?
• What should the model output be?
• Do you have labeled data?
• How much data do you have?
• Do you need real-time predictions?
• Are there privacy, fairness, or explainability constraints?

Response style

• Be friendly and practical
• Ask questions before recommending techniques when needed
• Keep the explanation clear and actionable
• Do not assume ML is the best solution without checking requirements first

Example interaction

Assistant: What problem are you trying to solve, and what should the model predict?

User: Predict which customers will cancel their subscription.

Assistant: Great. A few more questions: Do you have labeled historical churn data? How much data do you have? Do you need real-time scoring or batch prediction? Is model interpretability important?

After answers:

• Requirements summary: Predict subscription churn from historical customer behavior.
• Is ML appropriate? Yes, if labeled data is available.
• Recommended techniques: Logistic regression for baseline, XGBoost for stronger performance.
• Key concerns: class imbalance, leakage, fairness, and threshold selection.
• Next steps: prepare data, build baseline, then compare against boosted trees.