Readings: a check that survives an unattended run

Read these primary sources in alphabetical order. The selected scope is recorded in existing dated course factsheets; the questions below use only that scope.

Source 1 — Anthropic, Best practices for Claude Code

Read Give Claude a way to verify its work (tools/claude-code-best-practices; as_of: 2026-10-07; Source 1). The registered claim is the recommendation to give an agent a runnable test, build or comparison with a readable pass/fail signal. Product commands are not required for the studio.

Before the run, name that signal for your task. After the run, ask whether the agent could change the check itself. What must your verification preserve outside its scope?

Source 2 — GitHub Docs, Understanding GitHub Actions

Read Workflows (tools/github-actions; as_of: 2026-09-05; Source 1). The registered scope covers YAML workflows, event/manual/scheduled triggers and jobs. This is a reading about pipeline entry points; the studio needs no hosted workflow, token or paid service.

Draw an issue-triggered path to a proposed PR. Mark the stage that blocks an absent check. Does the presence of a job definition prove that it ran on the candidate you are reviewing? Use your own local check result to answer, rather than inferring success from the diagram.

Connect the readings to the studio

Use the notes to distinguish a model request from a controller gate. Bring one stage, one failure input and the observable block to studio. These two sources do not certify any workflow as safe or correct.