Skip to content

Week 3 self-check: Agent skills and the CLI

Learning objectives

Use these questions to check whether you can explain a skill boundary, test a near miss, and preserve command-line evidence before submitting A3. Answer guidance is available after you write your own record.

Minimal example

For a research-codebase skill, classify this request: “Implement the requested change and add a regression test.” It is a near miss because it asks for implementation, while the skill description is for a fact-only map before proposing a change.

Hands-on exercise

R1. Retrieval 1. Name the three fields that distinguish an intended trigger, your prediction, and what a CLI run actually showed.

R2. Retrieval 2. Name the command that checks your SKILL.md and trigger record, and state one claim that command cannot make.

  1. What did your repeated workflow do by hand, and which command was its verification?
  2. Which three requests should trigger your skill? Why is each one different in wording?
  3. Which request is the near miss? What would “stays quiet” look like in your CLI record?
  4. Which five fields let another reader identify a command result? Give your own values.
  5. What does skill_check.py prove, and what can it not prove?
  6. If observed differs from should_trigger, what must the revision section contain?

Verification

Run your own checker again and compare every predicted and observed value with the saved CLI evidence. A correct answer includes the command and revision, not only a final sentence.

Common mistakes

  • Confusing the intended trigger (should_trigger) with your prediction (predicted) or the observed record (observed).
  • Treating a checker pass as evidence that an agent used the skill.
  • Omitting the near miss or changing it until it triggers.