| description | Build features in small end-to-end slices, not big horizontal layers |
|---|---|
| alwaysApply | true |
Build a tiny end-to-end slice first, validate, then expand. AI agents tend to produce complete solutions in one leap without testing the critical path — massive review burden and rework ("slop").
- One vertical slice — touches all relevant layers for the simplest case (CLI + parser + db + test, etc.).
- Commit and validate before expanding — pre-commit runs format, lint, typecheck, tests on staged files (when AI/agent env vars trigger it).
- Lite-harden the slice —
harden-prlite mode after each slice (fix in working tree; commit when the user asks). - Expand outward from the working slice.
- Never build horizontal layers in isolation (all DB helpers before any CLI wiring, all docs before any working index path, etc.).
Feature layers and worked examples: tracer-bullets skill.
Before opening a PR: harden-pr full on origin/main...HEAD.