Skip to content

Week 3 readings: skills and command-line evidence

Learning objectives

Use a skill description as a testable selection boundary and read a command-line record as evidence with a revision, command, result, and verification step. The format's progressive-disclosure claim is scoped by (tools/agent-skills; as_of: 2026-10-07; Source 1).

Minimal example

Read the three course-authored skill examples in this order:

  1. Source 1 — research-codebase/SKILL.md and trigger-test.yaml: a read-only repository map, with three requests that should trigger and an implementation request that should stay quiet. Scope: description wording, source locations, and the distinction between predicted and observed selection.
  2. Source 2 — plan-change/SKILL.md and trigger-test.yaml: a bounded plan before implementation. Scope: acceptance criteria, allowed files, and a near miss that asks only for current facts.
  3. Source 3 — review-diff/SKILL.md and trigger-test.yaml: review evidence for an existing diff. Scope: reproducing findings, classifying them, and preserving the verification command.
  4. Source 4 — fixture/skill_check.py: the course checker. Scope: front matter, required sections, links inside the skill directory, four to twelve prompts, and the rule that a mismatch must be recorded and rechecked. It reads data; it does not run a model or select a skill.
  5. Source 5 — Agent Skills specification: the primary source named by (tools/agent-skills; as_of: 2026-10-07; Source 1). Scope: progressive disclosure from metadata to the activated body and then to bundled resources.

Open each source at the paths above. The trigger records are synthetic course examples; do not describe them as observations from a classmate, teacher, or model.

Hands-on exercise

Before the studio, annotate one source with two questions: which words in its description set the trigger boundary, and what output would prove a run happened? Bring the questions to the studio. No additional external reading is required for A3.

Verification

Use the direct command from the studio to check your own files. Keep the checker output separate from the agent transcript: they answer different questions.

Common mistakes

  • Treating a source file as an automatic host rule; the host's discovery and precedence must be observed for the agent you use.
  • Calling a synthetic trigger record a real model run.
  • Citing a skill card without naming which file and scope you read.