Course Engine Skill

Follow the six steps in order. Keep your narration terse.

1. Probe the session

Call the start MCP tool with the user's projectRoot (the working directory).

Course filter. When the invoking command names a course (a per-course start.md saying "filter to acc-claude-sdk"), drop every result.lessons entry whose course_name does not start with that plugin name. Without a filter, render every discovered course.

Render:

• Discovered courses: the filtered result.courses (plugin keys). If empty after filtering, say the named course is not enabled and stop.
• Lesson catalog: each filtered result.lessons entry with namespaced_slug, title, summary, and chapter_count, grouped by course_name. If empty, render result.warnings and stop.
• Warnings: each entry's kind + message when result.warnings is non-empty.

start never runs probes. Probes run in step 4, only for the picked lesson.

2. Output style

result.outputStyle is { active, recommended: "Concise", ok }. ok is true when active equals recommended.

• ok: true: one short confirmation, nothing more.

• ok: false: call AskUserQuestion with header: "Output style", question: "ACC works best with the Concise output style. Set it now?", options:

  • "Yes, set Concise (Recommended)"
  • "No, keep my current style"

  On yes, call setOutputStyle({ style: "Concise" }). Tell the learner the setting is written to ~/.claude/settings.json, and that /output-style Concise switches the current session immediately. On ok: false, surface errors verbatim, say the style was not changed, and continue. On no, continue.

The check is advisory. No tool refuses to run because of the active style.

3. Lesson selection

Ask the user which lesson to take. They refer to it by namespaced_slug (for example acc-claude-sdk@contract-hero/01-basic-agent).

Call selectLesson({ projectRoot, slug }).

If result.ok is false, surface result.errors verbatim and stop.

If result.ok is true:

• Render result.description verbatim when present. It is the lesson's description.md: what the learner is about to build, prerequisites, time.
• If result.workspaceCreated === true, say the workspace was seeded at result.workspacePath with the scaffold and the tests, without the solution files listed in result.workspaceStrippedFiles.
• If result.workspaceArchivedTo is set, say the host content changed, so an older workspace was archived and the lesson starts fresh at chapter 1.
• Render result.warnings when present, one line per { kind, message }. A state-corrupt warning means the previous state.json was unreadable. Name its archivedTo path when the warning carries one.

3a. First-run workspace setup

If result.firstRunSetup?.needsWorkspaceRoot === true, the user has never picked a workspace root on this machine. Run this one-time prompt before step 4:

• Surface result.firstRunSetup.defaultWorkspaceRoot (typically ~/workspace) as the recommended default.
• Call AskUserQuestion with header: "Workspace root", question: "Where should ACC put per-course project checkouts?", options:
  • "Use the default (~/workspace)" (Recommended)
  • "Pick a custom path": follow up in chat for the literal path; treat empty input as the default.
• Call configureWorkspace({ workspace_root: <chosen> }). On ok: true, say ~/.acc/config.json was written. On ok: false, surface the errors and stop.
• Refresh the workspace. The first selectLesson seeded the workspace against the default config, so its .env.acc-paths file points at default paths. Re-call selectLesson({ projectRoot, slug }) once after configureWorkspace returns ok: true. prepareWorkspace is idempotent on the host signature, so the workspace is reused and only the env file is rewritten. Use the refreshed result for the rest of the session.
• If result.firstRunSetup is absent, skip this step.

Paths declared by the course default to <workspace_root>/<manifest-default>. The learner can edit ~/.acc/config.json later, or call configureWorkspace({ course_paths: {...} }) for per-id overrides. Do not prompt for those unless the user asks.

3b. Resume or start over

result.resumed is true when state already held this lesson, the lesson is not complete, and the workspace was reused. The recorded artifacts are kept.

Only when result.resumed is true:

• Say the lesson resumes at chapter result.chapter_cursor + 1. chapter_cursor is 0-based.
• Say selectLesson({ projectRoot, slug, restart: true }) starts the lesson over at chapter 1.
• Call AskUserQuestion with header: "Resume", question: "This lesson is already in progress. Resume or start over?", options:
  • "Resume at chapter N" (Recommended): continue with this result.
  • "Start over": call selectLesson({ projectRoot, slug, restart: true }) once and use that result for the rest of the session.

When result.resumed is absent or false, ask nothing. The lesson starts at chapter 1.

4. Prerequisites

If result.prerequisites is a non-empty array, call runPreflightProbe({ probeId }) for each id.

• pass: true: move on.
• pass: false with no action: surface message verbatim. Tell the learner to fix it, then re-run the course's start command.
• pass: false with an action: surface the message and the action's command. Ask the learner whether to run the remediation. Only on an explicit yes, call runPreflightProbe({ probeId, remediate: true }). If the remediation also fails, stop the session.

Block step 5 until every prerequisite returns pass: true. Print one status line per probe.

If result.prerequisites is absent or empty, skip this step.

5. Personalization

If selectLesson returned a personalizationPrompts array, walk it: ask the user for each value, or accept the default. Collect the values into one object and call setPersonalization({ projectRoot, values }).

If the lesson has no personalization options, call setPersonalization({ projectRoot, values: {} }) to lock in the defaults.

Check result.ok. On ok: false, surface result.errors verbatim and ask the learner for the values again. Do not go to step 6 until setPersonalization returns ok: true.

6. Hand off to the conductor

Once selectLesson, every runPreflightProbe, and setPersonalization have succeeded, dispatch the course-conductor agent. The conductor reads state on its own and runs the chapter loop: per chapter it tells the learner what gets built, implements the code until the chapter's tests pass, writes the chapter's HTML artifact, calls verifyChapter, and pauses.

Say one sentence: the conductor takes over from here.