Skip to content

Playbooks

A playbook here is a repeatable procedure, not a tutorial. Each one assumes you know the engineering; it specifies only what changes when an agent does the typing: the exact prompts to issue, the order to issue them in, the points where a human must look before the work continues, and the failure modes specific to that workflow. You should be able to open a playbook mid-incident, copy the prompts, and run it.

Every playbook follows the same skeleton — Objective, When to use, Inputs, Step-by-step process with prompts inline, Human review points, Expected artifacts, Common failures, Recovery, Acceptance criteria — so once you've run one, you can run any of them.

Which playbook?

Click through: Idea to Spec · New Project · MVP Build · Add a Feature · Large Tasks · Production Debugging · Recovery · Legacy Refactor · Migration.

Two playbooks sit outside the tree because they apply to everything the agent produces: Testing AI Code and Reviewing AI PRs. Run them regardless of which branch you took.

All playbooks

PlaybookObjectiveRiskTypical duration
New ProjectStand up a repo with governing docs and quality gates before any featureLowHalf a day
Idea to SpecTurn a rough idea into ordered, session-sized, verifiable specsLow1–3 hours
MVP BuildShip an MVP as verified vertical slices on a walking skeletonMedium1–3 weeks
Add a FeatureAdd a feature to a production codebase without collateral damageMedium2 hours–2 days
Legacy RefactorRestructure legacy code with a behavior-change budget of zeroHighDays–weeks
Production DebuggingDiagnose and fix a live issue with the agent as investigatorHigh30 min–hours
MigrationMove stack, framework, or data store with verified equivalenceHighDays–weeks
Testing AI CodeVerify agent output against intent, not against the agent's own testsMedium1–2 hours per feature
Reviewing AI PRsReview high-volume AI diffs for behavior, not craftsmanshipMedium20–60 min per PR
Large TasksExecute work too big for one context window across sessionsMediumMultiple days
RecoveryGet back to a known-good state after an agent session goes wrongHigh1–4 hours

Builder/reviewer splits, parallel agents, and other multi-agent workflows are not playbooks here — they live in Multi-Agent Patterns.

  • Prompts vs Specs — why every playbook routes through a written spec instead of a chat message.
  • Responsibility Split — which decisions playbooks reserve for the human, and why.
  • Prompt Library — the standalone prompt collection the playbooks draw from.
  • Multi-Agent Patterns — builder/reviewer and parallel-agent workflows that compose with these playbooks.
  • Case Studies — the playbooks executed end-to-end on realistic projects, mistakes included.

A field manual for AI-native software engineering.