Session 02 · All class materials →
Lesson 2 — Codex as a Disciplined Teammate
Duration: 120 minutes
Included delivery track: Follow LAB_SETUP.md for the supplied files and local simulation. Required activities use these materials; connected deployments are optional extensions. Instructor debrief: answer key. Real OAuth, cloud releases and payments are simulated or discussed through evidence packets.
Deck: 02-codex-as-teammate
Lab: Codebase Detective
Learners will be able to
- give Codex a task brief with outcome, context, constraints, and completion bar;
- ask Codex to investigate an unfamiliar repository without granting broad authority;
- inspect a diff and a test result before accepting a proposed change; and
- record useful repository knowledge as a durable playbook instead of relying on chat history.
Instructor preparation
Slide 6 live demo: 15-minute instructor run sheet, exact prompt and fallback. Pre-open it before class.
Pre-open AGENTS.md — course AI instructions, requirement R1 in allocation-requirements.md, and issue G102 in github-issues.json, alongside the included sandbox. R1 and G102 are labels inside those files, not separate filenames. Distinguish the handout's staff report from the concierge issue: staff denial is correct. Explain that npm test checks fixture integrity, not release readiness. Scope model context; public answers do not replace learner reproduction.
Preflight clients before class using the small-model guide. Qwen3.8-27B is recommended; OpenCode, compatible Codex CLI or verified Claude Code routes are acceptable. Do not spend lesson time debugging a gateway or require premium inference. Prepare the GitHub kickoff, approved proposals and one brief comparison. The original investigation lab remains 35 minutes.
Agenda
| Time | Facilitation |
|---|---|
| 0–10 | Improve a vague task brief: goal, relevant files, scope and observable completion. |
| 10–25 | Small-model workshop: Qwen3.8-27B, client vs host vs model, vague/bounded task comparison, strengths/weaknesses and retry/spending limits. Use an approved route or labeled prepared outputs. |
| 25–40 | Codex/client demo: read project instructions, trace one path, verify citations and turn findings into a bounded brief. |
| 40–50 | Product kickoff part 1: confirm approval, create own public repo and initial issues. Unapproved pitches receive scope coaching; no product implementation yet. |
| 50–60 | Break. |
| 60–95 | Codebase Detective lab in pairs; inspect claims and run a check. Switch roles after planning. |
| 95–115 | Product kickoff part 2: create topic branch, sanitized README/proposal PR and focused review request. Coaches check model/host and budget records, capture the kickoff PR/head SHA, and explain how genuine history supports certification; finish pending human review after class. |
| 115–120 | Each learner identifies their first small product task, verification step and next review. |
Read the certification evidence standard. In the existing kickoff, show why a focused diff, linked issue and review response matter more than commit counts. Record the first checkpoint; do not introduce an extra exam or extend the agenda.
Teaching notes
Start the assessment agreement from the approved proposal: scope, environments and evidence routes for all rubric areas. Exact acceptance cases are finalized with the instructor during Session 04; no one needs to invent a full test suite at kickoff.
Use practical repository habits: provide relevant file paths and docs, split work into bounded tasks, keep repository instructions current, and run the project’s verification commands after a change. Modern Codex workflows support durable project context and tool-connected tasks; the course keeps human review and explicit boundaries at the center.
Formative coaching
Review the AI task brief and evidence log without numeric scores. A proposed code patch alone does not earn credit; the learner must explain which test or observable behavior checks it.
Product Studio connection — Kick off your own repository
Use up to five minutes of the existing closing/debrief time for a project next-step check; do not extend the 120-minute session. The build work happens as homework. Finish the approved-proposal PR, three initial issues, README and project instructions. Confirm your client/model/host and spending ceiling. Build only the first small approved slice; arrange human feedback on the kickoff PR.
Follow the milestone map and review checklist. Record support needs privately; practice lab quality criteria are coaching, not an exam or numeric completion gate.