What problem does it solve? Long-running agent-assisted projects accumulate scattered documentation, duplicated explanations, and stale implementation detail that degrade future agent effectiveness. This Skill classifies each new learning and routes it to the single authoritative surface—AGENTS.md, a canonical doc, an ADR, a Skill, a plan, research, tests, or code—so project knowledge stays lean and retrievable. ## Core Features & Use Cases - Knowledge Classification: Categorizes each insight by type (invariant, decision rationale, repeatable method, active state, external evidence, deterministic behavior) and assigns the narrowest correct destination. - Deduplication and Pruning: Searches for existing statements and contradictions, replaces duplicate explanations with links, and removes stale implementation detail from foundational docs. - Wiki Workflow Support: Manages the _index/ inbox-to-pages pipeline, respects wiki-generated markers, and runs pnpm wiki:test and pnpm docs:check after changes. - Use Case: After a user corrects the agent for the third time about a deployment constraint, use this Skill to fix the instance, identify the failure class, and record the durable prevention in AGENTS.md or an ADR instead of scattering notes across docs. ## Quick Start Use the project-knowledge-steward skill to decide where this retrospective finding belongs and update the canonical documentation accordingly.
