Tutorials
Hands-on walkthroughs of the practices we use to keep AI agents working on what was meant, not just what was measured. Each one is self-contained and needs only a coding agent, git and a shell.
-
Recognizing Extensional Drift in PracticeGreen tests, a clean lint score and a closed task can all coexist with work that missed the point. Three real cases from agent-driven development, and a short test for catching drift before it ships.
-
Plans That Record Decisions, Not GuessesTurn a plan file into the durable record of what you decided and what you rejected. One agent preps the open questions as facts; a fresh agent walks them with you, one at a time.
-
Iterative Planning Prompts You Can CopyFour copy-ready prompts that make an agent plan with you instead of for you: one question at a time, a research document built in rounds, interviewing a codebase first, and a handoff that gives goals and constraints rather than steps.
-
Put the Walls in Structure, Not PromptsA prompt that says “don’t” is not a boundary. Use sandboxes, blocking hooks, read-only checkouts and your agent’s permission rules so it can’t cross the lines that matter, and can explain why it stopped.
-
The Implementer Can’t Close Its Own WorkBuild a reviewer agent that can’t edit what it judges, a rework loop with a round limit, and a rule that every finding names the test that would have caught it.
-
Reviews That Can’t Quietly Drop FindingsAnchor an AI review to numbered markers so its coverage can be counted, not trusted. Includes a reviewer definition, a one-line self-check and a way to turn the report into a checklist.
-
Run a Debate Before You DecideA recipe for two agent sessions: one presents the question, one answers without seeing the plan, and every verdict must say what it rejected. Write your own answers first so the comparison is honest.
No published tutorials in this category yet — more are coming soon.