Ten focused rules. Each one a single Rule / Why / How-to-apply, mirroring Claude's memory conventions.
Quality rules that bind every step of the lifecycle.
Every step must measurably reduce uncertainty. Plans that don't are blocked at their gate.
AI detects, suggests, implements. The human decides scope, architecture, "done", and commits.
Every artifact records who proposed and who decided it. /approve is the only path to human-signed decisions.
Every decision is a click with concrete options. Free-text questions buried in prose are forbidden.
Read the required context (CLAUDE.md, plans, ADRs, code) before any change. Keep artifacts up to date.
Multi-target projects build & pass on all targets from commit #1. Retrofit is not an option.
Project & reference knowledge land in governance-eligible artifacts (ADR, Inventory, plan), never in free-form memory.
Core framework + domain framework + RAG. Layers compose; domains stay interchangeable.
Hierarchical §-numbering for plans — stable tree-structured IDs organized by chapters.
Plans match their type (CTX/EXP/PRD), size range, and minimum-substance requirements.
Every guideline uses memory-style frontmatter plus a Rule / Why / How to apply header, mirroring Claude's memory conventions. The R/W/H block is the minimum — sections are added only where they pay for themselves.
The format is canonical, defined by chevp-workflow/guidelines/README.md. When the canonical format changes, updates propagate via a workspace plan.