Choose a project small enough to understand, test and explain—not merely complete. The following is an original sample plan for a coding-ready learner, not a promise that every learner will finish in six sessions.
Explore the comparison in this briefing
Session
1. Frame the problem
- Main activity
- Decide categories, inputs and what success means
- Evidence saved
- One-page project brief and baseline
- Independent learner check
- Explain the problem without naming a tool
Compare each option against the same criteria. The complete table and qualifications remain in the briefing.
Example: a small hobby-item classifier
Use a deliberately simple, non-sensitive dataset of hobby objects or fictional items. The goal is to compare a straightforward rule with a small learned classifier, inspect errors and explain limitations. Keep the project local or in an approved private environment.
| Session | Main activity | Evidence saved | Independent learner check |
|---|---|---|---|
| 1. Frame the problem | Decide categories, inputs and what success means | One-page project brief and baseline | Explain the problem without naming a tool |
| 2. Inspect the data | Examine labels, missing examples and a simple rule-based baseline | Data description and baseline results | Explain one example and one ambiguous case |
| 3. Train and test | Build the first model with appropriate separation of training and evaluation data | Small runnable workflow and test plan | Explain why evaluation examples must be protected from training decisions |
| 4. Investigate errors | Examine mistakes and propose a limited improvement | Error log and revised hypothesis | Predict what a proposed change might affect |
| 5. Make it understandable | Add a minimal interface or clear demonstration | Instructions, limitations and contribution record | Modify a small requirement without the mentor taking over |
| 6. Demonstrate and hand over | Re-run the workflow and discuss failures | Final files, short README and learning review | Explain, change and debug a small component |
For model-based work, evaluation discipline is not decorative. Scikit-learn’s documentation explains that leakage between training and test information can produce overly optimistic performance estimates; it recommends splitting appropriately and avoiding learning preprocessing or model choices from the test data. Treat a tiny educational dataset as a teaching exercise, not proof of real-world accuracy. (scikit-learn.org)
What if the teenager specifically wants a generative-AI application?
Keep the same six stages, but change the evidence:
- Use a small approved information collection, not the family’s real correspondence.
- Write expected answers and unacceptable behaviours before optimising prompts.
- Include questions the application should decline or label uncertain.
- Compare outputs with the underlying source material.
- Record errors and changes rather than showing only successful examples.
- Start with a private demonstration, without autonomous purchasing, messaging or account access.
NIST’s generative-AI risk profile identifies risks including confidently false outputs, privacy exposure, information-security failures and inappropriate human reliance. These are reasons to teach testing and boundaries alongside building—not reasons to promise that one checklist makes a prototype safe. (nvlpubs.nist.gov)