superdev-self-learning

name

superdev-self-learning

description

The meta-skill that makes superdev self-improving. Captures user-frustration signals (explicit corrections, code reverts, regression-verifier failures, design-drift above 1%, re-prompts with frustration markers) and writes structured **feedback** memory entries to `.claude/memory/superdev-learned/`. Future orchestrator dispatches read this memory first so the system learns what NOT to do in this project, what skills to call (or never call) in which context, and which defaults to pre-apply. Project-scoped only; never global. Triggered automatically by the `UserPromptSubmit` hook on frustration patterns and by `SubagentStop` after fix-applier or design-fidelity-auditor.

Superdev Self-Learning — the meta-loop

The skill that makes superdev get smarter every time it fails. When you correct it, when you revert its code, when a verifier catches a regression, when an audit fails — the system learns the lesson, writes it to project memory, and primes every future agent dispatch with that lesson.

This is what turns "a plugin" into "the ultimate superdev team".

The Iron Law

EVERY FRUSTRATION SIGNAL BECOMES A FEEDBACK MEMORY ENTRY.
EVERY ORCHESTRATOR DISPATCH READS `.claude/memory/superdev-learned/` FIRST.
THE SAME MISTAKE MUST NEVER BE MADE TWICE IN THE SAME PROJECT.

When this skill is invoked

Automatically (via hooks — see below):

• UserPromptSubmit hook detects frustration keywords ("no", "wrong", "stop", "that's not what I asked", "I already told you", "for the third time") AND there were recent superdev edits in this session
• SubagentStop after fix-applier succeeds (extracts the LESSON: line)
• SubagentStop after design-fidelity-auditor reports drift (extracts drift-cause patterns)
• SubagentStop after regression-verifier REJECTs (the rejection IS a lesson)