Reflect

**Verdict:** A strong, opinionated reflection and learning skill that’s well-suited for serious multi-session, multi-agent setups, especially in repos that already use `.claude/`-style agent files and git. It introduces real, durable self-improvement while maintaining a conservative safety posture. **What it does well** - Turns user corrections and successful patterns into structured proposals for updating agent files or creating new skills. - Uses confidence-based signal detection (high/medium/low) and routes learnings appropriately (agents, skills, project memory, global memory). - Enforces human-in-the-loop review: shows diffs, supports selective application, and optionally commits via git with descriptive messages. - Maintains state and metrics so you can track accepted vs. rejected learnings and reflection activity over time. - Integrates with hooks (e.g., PreCompact) and auto-reflection modes for teams that want continuous, low-friction improvement. **Main risks and limitations** - **Setup complexity:** Requires understanding of the file layout (`.claude/`, state dirs, hooks) and possibly git; non-technical users may struggle without guidance. - **Signal quality:** Pattern-based detection can misclassify or miss learnings; the system is only as good as the input signals and user review discipline. - **Overfitting / clutter:** Aggressively encoding every minor preference into agent files may create bloated, conflicting rules if users don’t curate proposals carefully. - **Environment assumptions:** Best fits workflows that already mirror the documented structure (CLAUDE.md, agent files, git repo); outside that, some benefits may be reduced. **Best suited for** - Engineering or data teams running long-lived projects that want their agents to learn from real feedback rather than just static prompts. - Power users maintaining multiple specialized agents and skills in a codebase, with comfort using git and repo-level configuration. - Organizations that value auditability and safe iteration—wanting clear history of agent behavior changes and easy rollback. **Less ideal for** - Casual or short-lived chats where maintaining reflection state and agent files is overkill. - Users without access to filesystem-level operations or git, or who prefer very lightweight, one-off interactions over systematic improvement.