session-intelligence-harvester

The session-intelligence-harvester is an essential meta-skill for Spec-Driven Development (SDD) that transforms ephemeral session artifacts into durable organizational assets. Instead of allowing fixes for format drift, orchestration errors, or missing context to remain isolated in a single session history, this skill systematically routes identified lessons into the appropriate RII components, including CLAUDE.md, agent constitutions, and skill specifications. By bridging the gap between "getting it working" and "getting it encoded," it ensures that the agent workforce becomes smarter over time, effectively compounding the intelligence of the development process.

• Automatically identifies candidates for harvesting based on failure modes, repeated pattern corrections, and multi-file updates.

• Implements a structured 5-step workflow for session analysis, routing, reading, editing, and validation.

• Integrates with the repository's .claude/ directory, updating specialized agents, skills, and command protocols in real-time.

• Enforces strict routing logic to ensure learnings trigger at the correct stage, such as adding context-gathering gaps to CLAUDE.md or new validation steps to command orchestration files.

• Supports manual triggering when users identify significant project discoveries or documentation gaps that require immediate codification.

• Operates best when the user identifies recurrent failure modes or discovers new convergence patterns across multiple files.

• Expects the user to confirm the analysis of what was wrong versus what is now correct to ensure proper knowledge encoding.

• Prioritizes direct action by editing target files to apply updates immediately rather than merely suggesting proposals.

• Requires existing RII structures to function correctly; ensure the target repository follows the standard .claude/ folder hierarchy.

• Encourages users to classify learnings into specific types like pedagogical issues, agent convergence patterns, or engineering skills to maintain clean knowledge architecture.