Skip to content

A3: Turn a manual workflow into a skill

At a glance

  • Task: perform a small workflow by hand, encode it as one skill, and test its trigger and near miss.
  • Week: 3
  • Weight: 2.5 points
  • Due: Sunday 23:59 in the listed week
  • Hand in: the skill file, manual-work evidence, two trigger trials and report.json.
  • Done when: the intended request triggers and a similar request stays quiet, with outputs you can inspect.
  • Time: published after a timed pilot; record your prediction and actual time.

Objectives

Write a precise, bounded skill and evaluate when it should activate (LO2).

Prerequisites

Follow the safety practices in syllabus §9:

  • Run the agent in a dedicated directory or container.
  • Keep residency, employer and cloud credentials out of that environment.
  • Keep keys out of repositories, prompts, screenshots and logs.
  • Run injection demonstrations offline against the local machine only.

Use the mature coding-agent seat supplied for the course. Do not run an unknown generated command on a credential-bearing host.

Tasks

  1. Run by hand. Complete the workflow without a skill. Save the request, inputs, steps, expected output and verification command.
  2. Generate. Ask the agent to draft one SKILL.md with a name, trigger, non-trigger, steps, limits and verification.
  3. Verify the trigger. Use one request that should activate it and inspect the resulting trace or activation marker.
  4. Verify the near miss. Use a similar request that should remain quiet. Record what you observed, including a false activation if it occurs.
  5. Revise. After feedback from a peer or the teaching team during the assignment week, tighten one trigger or limit and rerun both cases. Record who gave it and the date; if no feedback was available, record your self-review date and explain the choice in the decision line.

Generation is the proposed skill; verification is the manual baseline and the two controlled requests. Do not count a plausible activation as proof without an observable marker.

Deliverables

Include the skill, manual evidence, both trial records and report.json. Run:

python practice/a3/check.py path/to/submission

Grading

Criteria are a reproducible manual baseline; an unambiguous trigger; silence on the near miss; bounded instructions; and evidence for the revision. The public check is a consistency check, not a model judge.

AI use rules

Green: drafting skill prose and test requests. Yellow: suggesting trigger boundaries, followed by your own runs. Red: deciding whether activation occurred and choosing the final decision. Attach one AI-use note ID.

Estimated time

Published after a timed pilot. Record predicted and actual time and ask for support if the agent's activation signal is not observable.

Work record outline

A3 report
manual_evidence: [TODO]
skill_path: TODO
trials: [{case: trigger|near_miss, request: TODO, expected_trigger: true|false, actual_trigger: true|false, evidence: [TODO]}]
revision: TODO
explanation: TODO
ai_use_note: TODO
decision: accept | narrow | recheck | stop
because: evidence ID(s)
next: TODO