The part most reference material skips: how people actually put these ideas together. Each playbook is a sequence you can follow, wired to the terms it depends on.
Ship a feature with an agent, safely
You want an agent to implement something real, and you want to review the result rather than the process.
Write the spec first. State behaviour and acceptance criteria in the ticket. Anything you leave implicit is a coin flip.
Grill the spec. Ask the agent to interrogate the plan and list ambiguities before writing code. Fix the spec, not the code.
Let it build in a branch. Give it the repo, a test command and permission to iterate. Cap iterations and keep it inside the workspace.
Gate on automated checks. Types, tests and lint must pass before you look. Do not spend your attention on work a machine can reject.
Review the diff, not the transcript. Read the change against the spec. Watch specifically for tests being weakened to make them pass.
Prompt you can copy
Read the spec in this ticket. Before writing code, list every ambiguity, missing edge case and unstated assumption, and ask me about each. Then implement on a branch, run the test suite, and open a PR that references the spec.
Build a RAG assistant over internal documents
People ask questions whose answers live in documents the model has never seen.
Debug a hallucinating assistant
Users report confident answers that are simply wrong.
Cut inference cost without losing quality
The pilot works and the bill is now a problem.
Run an overnight migration factory
You have hundreds of mechanical changes that can be verified automatically.
Verify a prompt change before shipping it
Someone wants to edit a prompt that drives production behaviour.
Choose a model for a task
You are starting a feature and need to pick a model without a three-week bake-off.
Write an AGENTS.md that actually helps
Agents keep making the same project-specific mistake.
Roll out an internal AI tool people actually use
You have built something good and adoption has stalled in the low teens.
Red team an agent before it goes live
The agent has tools, credentials or access to other people’s data.
Rescue a session that has gone stale
The agent is revisiting rejected ideas and contradicting decisions made an hour ago.
Build a structured extraction pipeline
You need to turn messy documents into rows your code can use.