instinct-system

To detect project conventions, the instinct system scans related files to form confidence-scored hypotheses about recurring patterns, adjusting scores based on observed confirmations or contradictions. It actively seeks evidence across handlers, naming, architecture, testing, and API design.

Confidence scoring for code pattern detection works by creating instincts at low confidence and raising or lowering them through an observe-hypothesize-confirm lifecycle. Well-supported instincts with 0.9 confidence can be promoted to permanent MEMORY.md records.

To stop the assistant from guessing conventions, use the instinct system to capture and validate recurring project-specific patterns. It prevents one-off observations from becoming incorrect rules by learning conventions through active confirmation across related files.

When scaffolding features, the instinct system detects conventions like handlers returning Result<T> across the codebase, confirms them across related files, then applies the convention automatically once confidence reaches the follow threshold.

Yes, you can export and share learned project conventions between codebases. The instinct system supports export and import functionality with decay for cautious sharing, ensuring conventions are validated rather than blindly applied across projects.

Confirmed project convention records are maintained in .claude/instincts.md. The system maintains per-project records, flags stale items, removes dead instincts, and promotes well-supported instincts to MEMORY.md at 0.9 confidence.