Today’s outcome

• Write an executable AI task brief
• Investigate an unfamiliar codebase safely
• Review a proposed diff and its tests
• Preserve useful context as files and playbooks

Treat Codex like a capable new teammate

Give it:

• desired outcome
• relevant files and sources
• scope and safety boundaries
• completion bar

Then review its work like you would any teammate’s.

A strong task brief

Find why staff cannot send an inquiry update. Read these paths. Do not change files. Return: request path, cited evidence, likely hypotheses, open questions, and the smallest verification command.

Specific task. Bounded authority. Reviewable result.

Modern Codex workflow

Explore → plan → review → apply → verify

For unfamiliar or consequential work, ask for a plan first. Use repository instructions and current source files as durable context.

Demo · 15-minute read-only investigation

1. Open AGENTS.md, requirement R1, and issue G102
2. Ask for a bounded investigation plan
3. Trace UI → server → canSend
4. Verify the references and run npm test
5. Save the result, limits and open question

Open the instructor demo run sheet

What Codex should say

• “Here is the evidence I found.”
• “This is an inference.”
• “This requirement is ambiguous.”
• “I need approval before changing…”

It should not pretend to know missing business rules.

Small models, serious work

Recommended: Qwen3.8-27B · summer 2026 release

Local if your hardware supports it; capped hosting otherwise.

Model ≠ hosting provider ≠ coding client ≠ authorized tools.

Verify one file → edit → check loop before starting.

Spend on progress, not retries

Good fit: bounded features, tests, docs and structured data.

Watch for invented rules, lost context and repeated tool errors.

Give relevant files, one outcome and an observable check.

After two failed attempts: narrow the task and ask for help.

Your route and your budget

OpenCode + OpenRouter is a documented hosted route.

Codex custom/local endpoints need a compatibility check.

Claude Code + Qwen is not guaranteed; verify or use another client.

Suggested project target $10; agreed cap, no automatic premium fallback.

No subscription or new hardware purchase is required.

Lab: Codebase Detective

Mission: investigate, do not rush a fix.

Deliver: task brief, verified repository map, tool result, and open question.

Diff review questions

• Does the change match the acceptance criterion?
• Is its scope small and reversible?
• Would the test fail for the actual regression?
• Did it leak data, broaden permissions, or hide a failure?

Durable context beats chat memory

Keep:

• repository instructions
• test commands
• architectural notes
• approved playbooks
• evaluation cases

What takes an agent further?

Context: current requirements, files and examples.

Tools: actions the client can execute.

Reusable workflow: instructions and scripts for repeated jobs.

Feedback: tests, observed state and corrections.

Tools, MCP, and reusable skills

MCP connects an agent to external tools and context.

A skill packages a repeatable workflow with instructions and optional scripts/resources.

Account access and permissions determine what can run.

Let it execute an authorized task

Ask Codex to inspect, implement, run checks, diagnose failures, and iterate within the learner copy.

Give a clear stopping condition and review the result.

Use extra reasoning or parallel work when it improves measured outcomes.

Exit ticket

Write a task brief asking Codex to find, not fix, a suspected regression.

Real history, not a commit-count target

Plan → working slice → tests and fixes → reviewed release.

Keep focused branches, PRs, feedback and actual revisions.

Explain one issue-to-release chain and the diff it contains.

Squash merges are fine when evidence remains traceable.

No manufactured milestones. No surprise Git quiz.

Agree the destination before judging the result

Session 02: approved scope, environments and skill mappings.

Session 04: exact inputs, expected results and recovery rules.

Keep the agreement version with your project evidence.

Different products. Same skills. No surprise final-day bar.

Kickoff: your product repository

Approved proposal → your own public repo → three initial issues.

Create a topic branch; open a focused README/proposal PR.

Review the diff, invite feedback and record the next small task.

Code stays in your repo. Practice packets go to lab-submissions.