What problem does it solve?
It helps you reliably detect recurring project conventions by treating early observations as hypotheses, scoring confidence as evidence accumulates, and only promoting confirmed insights into permanent guidance.
Core Features & Use Cases
- Confidence-scored hypothesis lifecycle: Start instincts at 0.3, adjust confidence through confirm/disconfirm evidence, and promote only when confidence reaches 0.9.
- Project-scoped memory: Keep instincts per-project in .claude/instincts.md and promote them into MEMORY.md once confirmed to avoid cross-project contamination.
- Observe–Hypothesize–Confirm workflow: Actively seek evidence in 2–3 related areas, track seen counts and last-updated dates, and discard stale or contradicted patterns.
- Confidence-based behavior control: Ignore low-confidence items, mention or follow at higher confidence, and request user approval before permanence.
- Import/export discipline: Export at 0.7+ and import with decay so local verification remains required.
Quick Start
Load this skill at the start of a session to read .claude/instincts.md and apply any 0.7+ instincts, then when you notice a recurring pattern form a new instinct at 0.3 and check 2–3 related files to confirm or disconfirm it.