Intermediate Worksheet Pack - Paired Engineering with AI

Pack intent

These worksheets are for practitioners who can already use AI productively but still need more discipline around ambiguity, verification depth, and system effects.

Facilitation defaults

• preferred format: pairs or triads
• recommended timebox: 20-35 min per exercise plus 10-15 min debrief
• recommended AI stance: choose mode deliberately; do not default everything to delivery mode

Worksheet 1. Flaky Test or Race Condition

Canonical scenario:

• Intermediate Exercise 1 - Flaky Test or Race Condition

Use when:

• teams overuse timeout increases and sleeps
• the facilitator wants to expose the difference between stabilizing and understanding

Suggested format:

• 20 min working time
• 10 min debrief
• pairs

Facilitator setup:

• keep the quick-fix temptation visible
• ask learners to name what evidence would separate flaky test from real race

Learner handout prompt:

Use AI to compare multiple explanations for the intermittent failure, but do not start with a blind timeout increase. Define what evidence would distinguish test fragility from system behavior.

What good looks like:

• the learner rejects shallow stabilization as proof
• the investigation path includes instrumentation or timing evidence
• uncertainty remains visible until evidence improves

Likely traps:

• quick-fix bias
• confusing CI rarity with low importance
• missing observability gaps

Debrief prompts:

• what evidence would justify a timeout change
• what did AI help you consider
• what still required engineering judgment

Worksheet 2. Query and Cache Tradeoff Review

Canonical scenario:

• Intermediate Exercise 2 - Query and Cache Tradeoff Review

Use when:

• teams tend to accept “just cache it” as universal advice
• the facilitator wants to reinforce tradeoff reasoning

Suggested format:

• 25 min working time
• 10 min debrief
• pairs or triads

Facilitator setup:

• keep freshness and support cost visible
• ask learners to define what data they wish they had

Learner handout prompt:

Use AI to compare query optimization, partial caching, and full payload caching. Then decide what evidence you still need before recommending any one path.

What good looks like:

• clear tradeoff comparison
• measurement needs are explicit
• cache invalidation and freshness are treated as real design issues

Likely traps:

• assuming performance advice is context-free
• optimizing one metric at the expense of trust or operability
• accepting AI benchmarks as local truth

Debrief prompts:

• what made one option look attractive at first
• what hidden cost mattered most
• what would you measure next

Worksheet 3. Policy Change With Hidden Edge Cases

Canonical scenario:

• Intermediate Exercise 3 - Policy Change With Hidden Edge Cases

Use when:

• AI outputs look polished but domain semantics are easy to miss
• the facilitator wants to pressure-test semantic verification

Suggested format:

• 30 min working time
• 10 min debrief
• small mixed-role groups if possible

Facilitator setup:

• require participants to tie rules back to policy text
• emphasize counterexamples

Learner handout prompt:

Critique the AI-generated decision logic instead of admiring it. Surface missing questions, edge cases, and places where the policy still needs clarification before implementation.

What good looks like:

• unresolved questions remain visible
• examples and counterexamples are used to test the logic
• the group resists the temptation to call the decision tree “done”

Likely traps:

• confusing completeness of format with completeness of policy
• hiding ambiguity inside elegant structure
• skipping stakeholder escalation

Debrief prompts:

• what looked convincing but unsafe
• which unresolved question most affected implementation
• what counterexample stressed the logic hardest

Worksheet 4. Contract Change Rollout Review

Canonical scenario:

• Intermediate Exercise 4 - Contract Change Rollout Review

Use when:

• teams need practice seeing past “optional means safe”
• cross-service coordination is a common local weakness

Suggested format:

• 30 min working time
• 10 min debrief
• triads

Facilitator setup:

• keep producer and consumer perspectives visible
• ask learners to distinguish protocol compatibility from rollout safety

Learner handout prompt:

Use AI to help evaluate the rollout plan for the contract change, then define what still needs real design review and cross-team verification before the change should ship.

What good looks like:

• strict consumers are treated as first-class risks
• rollback and observability are included
• “compatible” is defined carefully rather than assumed

Likely traps:

• trusting optional-field logic too quickly
• ignoring downstream assumptions
• skipping rollout sequencing

Debrief prompts:

• what made the AI answer sound safer than it was
• which consumer created the real risk
• how would you stage the rollout

Worksheet 5. Partial Incident Reconstruction

Canonical scenario:

• Intermediate Exercise 5 - Partial Incident Reconstruction

Use when:

• learners need practice in disciplined uncertainty
• the facilitator wants an incident exercise without pretending the root cause is knowable from thin evidence

Suggested format:

• 35 min working time
• 15 min debrief
• triads

Facilitator setup:

• ask for explicit confidence language
• require both internal reasoning and stakeholder-safe communication

Learner handout prompt:

Use AI to organize the incident, but do not let it decide the answer. Produce a hypothesis map, immediate containment options, a verification plan, and a stakeholder update that is accurate under uncertainty.

What good looks like:

• the update is useful without overclaiming
• hypotheses stay separate from facts
• containment and evidence collection are both visible

Likely traps:

• neat narrative bias
• overconfident stakeholder language
• collapsing multiple mechanisms into one story

Debrief prompts:

• which part of the AI output was most tempting to trust
• what uncertainty had to stay in the stakeholder message
• what evidence would narrow the hypothesis set