Learning

**Judgement:** A well-scoped, generally high-value meta-skill for making other skills behave like a personalized tutor. Its core idea—learning how the user learns and encoding that into compact preferences—is strong and broadly applicable, though the implementation appears lightweight and heavily reliant on the model’s internal reasoning. **What it does well** - Gradually builds a user-specific learning profile (style, preferred formats, tools, and “never” patterns) from interaction evidence. - Encourages confirmation only after “2+ consistent signals,” which helps avoid reacting to a single noisy interaction. - Keeps entries ultra-compact, which is important for skills that will be used over many turns or long sessions. - Works across learning contexts (school, work, hobbies) without hard-coding any domain. **Risks / Limitations** - **Opaque adaptation logic:** The actual update logic is mostly implied (“auto-evolves; edit sections as you learn”), so effectiveness depends on how consistently the host system or model interprets and applies these instructions. - **Overfitting to sparse signals:** With limited history, it may infer preferences too early or from coincidental successes unless the “2+ signals” rule is followed rigorously. - **Misclassification of preferences:** The model might misread politeness or time pressure as preference signals (e.g., assuming “short answer” is always desired), which could reduce explanation quality. - **No explicit safeguards against reinforcing bad habits:** If a user prefers overly shallow explanations, the system might converge to that pattern without balancing for actual learning efficacy. **Recommended scenarios** - Long-term tutoring or coaching bots where the same user returns frequently and the system can accumulate reliable preference data. - Study helpers and exam-prep assistants that need to vary explanation style between users (e.g., visual vs. procedural learners). - Workplace assistants that train or support staff and must adapt explanations to different roles and seniority levels. - Any multi-step teaching flow (courses, guided projects, onboarding sequences) where consistency with the user’s favored explanation format is important. **When to be cautious** - One-off, single-turn interactions where there isn’t enough evidence to infer stable preferences. - High-stakes educational contexts that require pedagogical rigor; the skill should be combined with content-quality checks rather than used as the only adaptation mechanism. - Deployments where transparency is required; implementers may want to surface the inferred preferences to users so they can review or override them.