# Practice and Question Routing

## Diagnostic prompts by gap

| Gap | Primary exercise | Follow-up challenge |
|---|---|---|
| Product discovery | reconstruct a recent workflow and baseline | prove the request is not just a solution idea |
| AI fit | compare process, rule, ML, LLM, and human | define stop criteria for the AI experiment |
| PRD | write behavior contract and failure states | trace requirements to evaluation and operations |
| RAG | diagnose a wrong answer layer by layer | design permission and freshness tests |
| Agent | choose autonomy level and tools | handle loop, duplicate action, and rollback |
| Evaluation | build a stratified 50-case starter set | calibrate judge and analyze severe tails |
| Project | create phase-gated WBS and RAID | recover a two-week delay through explicit tradeoffs |
| Safety | threat-model data and actions | respond to a cross-tenant leak in 30 minutes |
| Business | calculate value and successful-task cost | choose pricing and test a moat claim |
| Leadership | write a decision memo with dissent | stop a politically supported project using evidence |

## Answer structure

Train answers to follow:

1. Clarify user, goal, scope, risk, and decision.
2. State current baseline and available evidence.
3. Present options, including simpler alternatives.
4. Make a choice and explain tradeoffs.
5. Define metrics, failure modes, and safeguards.
6. Explain experiment, rollout, ownership, and stop/rollback.

Avoid forcing this into a rigid script when the question is narrow; preserve the decision logic.

## Follow-up ladder

Use increasingly difficult follow-ups:

- What evidence supports that?
- What is the non-AI baseline?
- Which assumption is most likely false?
- What happens in the worst 1% of cases?
- Who is authorized to make or reverse the decision?
- How will you test this before building everything?
- What would make you stop?
- How do cost and operations change at 100× scale?
- How would a strong competitor copy this?
- What mechanism prevents the same mistake next quarter?

## Practice modes

### Rapid oral drill

Give 2 minutes to structure and 3 minutes to answer. Score clarity, breadth, evidence, and closure. Repeat weak questions after 2–3 days.

### Written case

Give 30–60 minutes. Require one-page decision plus product/evaluation/project appendix. Score the artifact before giving a model answer.

### Artifact red-team

Review an existing PRD or plan. Identify missing evidence, invalid assumptions, untested failure modes, unclear ownership, and absent rollback. Require the learner to revise.

### Simulation

Introduce a changed constraint: vendor price increase, privacy incident, delayed data, model regression, executive deadline, or low adoption. Evaluate adaptation without hiding tradeoffs.

## Portfolio evidence

Require artifacts with sanitized context, personal contribution, decisions, alternatives, metrics, failures, learning, and outcome. Distinguish real work from synthetic practice. Never fabricate impact numbers or employer details.

## Feedback style

Lead with the highest-impact gap. Cite the exact answer or artifact evidence. Explain why it matters in real work. Give a concrete revision task and success rubric. Limit each cycle to a few priorities so the learner can improve.