skill-learner

Skill Learner

Turn user corrections into persistent knowledge that survives across sessions.

NEVER

• NEVER save a vague correction ("do it differently") without specifying the exact alternative — vague corrections create false confidence and are worse than no correction
• NEVER create duplicates — check INDEX.md first; merge if same skill+issue exists
• NEVER save one-time preferences as universal rules ("I wanted blue" ≠ "always use blue") — ask about scope when intent is ambiguous
• NEVER exceed 50 active corrections — consumption degrades; archive minor corrections older than 90 days to archive/
• NEVER write corrections in a language different from the user's — nuance dies in translation
• NEVER skip "When to apply" — a scopeless correction gets over-applied to contexts where it causes new problems
• NEVER save corrections that bypass safety ("don't validate input", "skip auth checks")
• NEVER assume old corrections are still valid — skill updates silently invalidate them; verify before applying corrections older than 90 days
• NEVER save without passing the cold-reader test: "Can a different agent in a different session act on this without any conversation context?" If no, rewrite before saving

The Correction Paradox

More corrections ≠ better behavior. Over-correcting creates rigidity — an agent drowning in 50 corrections becomes cautious and slow, second-guessing every decision. The goal is not to capture every complaint, but to capture the corrections that will prevent the most damage across the most future sessions.

Before saving, ask yourself:

• Frequency: Will this situation come up again? (One-off = don't save)
• Blast radius: If uncorrected, how bad is the impact? (Minor polish = maybe don't save)
• Generalizability: Does this apply beyond this specific conversation? (Too narrow = don't save)

If all three are low, tell the user you've noted it but saving a correction would add noise. They can override you.

Quick Classification