Tutor notes. For the person running the sessions.
Session 2: tutor notes
Objective: they internalise that output quality is mostly input quality,
gets the context-hygiene habits (fuel gauge, /clear, two-corrections rule,
save/restore as ritual), and, non-negotiable, one properly planned
increment live before they go home: backlog from the wrongs list, real
Definition of Done, the full loop run once together. The loop is the
deliverable; their solo run in this session just repeats it.
Timings (2h)
| Time | Segment |
|---|---|
| 0:00 | Homework review, their wrongs list drives this session's examples and becomes the backlog |
| 0:10 | /context-restore + context window & rot (/context live) |
| 0:25 | Prompt anatomy, rewrite 2 of them failed prompts together |
| 0:45 | Plan mode walkthrough (brief, /autoplan is minutes away) |
| 0:55 | Backlog + sprint board + DoD; /spec the slice; /autoplan |
| 1:20 | Build → /qa → /setup-deploy (+.env pattern) → /land-and-deploy → /canary |
| 1:55 | /context-save, this session's solo loop agreed |
Tight session, the loop must complete. Cut prompt-anatomy time before cutting the deploy; a live ugly slice beats a beautiful local one (say that sentence to them, it's the whole methodology).
Delivery notes
- Use their prompts as the teaching material. Pull up what they actually typed in Session 1. Rewriting their own vague prompt is worth ten invented examples, and the rewrite drops straight onto the backlog as a slice.
- The two-corrections rule is straight from Anthropic's own best-practices guidance. Present it as a rule with a reason attached (polluted context), and in this session you can give the reason, because context rot is on the menu.
- The backlog beat: the wrongs list becoming a backlog is the session's narrative hinge, "last session's failures are this session's plan" is agile in one sentence. Let them feel the promotion.
- Scope defence starts now. When they want a third slice in the sprint, the answer is the backlog, not "no". DoD enforcement is the other hill: an increment that skipped /qa or whose canary wasn't green does not move to Done, however charming its author.
/autoplan: meeting the delegate first
They meet the automated review board before they're ever sat on one, the reverse of how professionals usually earn delegation, and that's deliberate: slice-sized stakes are exactly what auto-adjudication is for. The line to plant (it pays off on Session 5): "there are four reviewers inside this thing. For the big project, you'll sit in every one of those chairs yourself." Have them read the auto-decisions scroll past and stop on one: "would you have ruled the same way?" If they override one with a reason, that's the whole course working. Note /autoplan is token-hungry: run it early: this is one of the heavier sessions.
Deploy session prep (before the session)
- Netlify (or your /setup-deploy target) pre-authorised for them GitHub, as in Session 1, the first /land-and-deploy must not die in an auth maze.
- Know what /canary actually watches on their stack so you can narrate it.
- Rehearse the .env demo: commit a fake key to a scratch repo, show it in history after deletion. Thirty seconds; lands forever.
Sticking points
- They writes essays as prompts. Longer isn't the lesson, decided is. Show that 4 crisp lines beat 3 rambling paragraphs.
- Plan mode feels like bureaucracy. Concede it! It IS overhead, teach the calibration (one-sentence change → skip planning). And in this session the payoff is immediate: /autoplan is plan mode with teeth.
- They want to spec the whole product in this session. "Spec the slice, not the system", the system spec is the backlog, and it's allowed to be vague. (The full-system version of /spec gets its own session later in the course.)
- Over-rotation on
/clear: some people start clearing compulsively and lose useful context mid-task. The judgement call: mid-problem with valuable history → keep going; new task → clear. - The DoD feels heavy for a one-session slice. It is the slice. "Done means done" is the one Scrum hill worth dying on in this stretch.
PM war story slot
A project that built exactly what the spec said and not what anyone wanted. Requirements as guesses vs requirements as decisions.