Agent Skills — Feedback Loop

Use this skill to wire a learnings loop into another skill so it gets better with use, without rewriting SKILL.md automatically.

The loop has four moving parts:

1. learnings.md — raw, append-only, committed. Shared working memory across machines; created on first append via append_learning.py, not seeded empty.
2. learnings.consolidated.md — pruned, dated, committed. Portfolio-grade institutional memory; seeded at wiring time.
3. learnings.local.md — machine-specific notes, gitignored. Use for one-operator-on-one-machine context that should not propagate.
4. scripts/append_learning.py + scripts/consolidate.py — keep raw entries well-shaped and promote durable ones to consolidated.

The name is borrowed from the 2026 "learnings loop" pattern (MindStudio) and Anthropic's own skill-authoring guidance to ask Claude to self-reflect on what went wrong and fold it back into reusable context. The mechanism deliberately does not match MindStudio's: MindStudio's loop has the model rewrite the skill's persistent instructions directly from user corrections — that is the exact auto-rewrite this design forbids (see Anti-Patterns). This skill keeps the same "accumulate corrections across sessions" shape but routes it through append-only raw entries, human-reviewed consolidation, and an eval-gated promotion step before anything touches skill logic — mapped onto this repo's existing 4-type memory schema (see agents-memory).

Quick Reference