I kept correcting Claude Code the same way across sessions:
"Don't catch that error here—let it bubble up so the global handler logs it with the request context."
"We don't use feature flags for this repo. The deploy pipeline handles rollouts."
"Stop creating new API endpoints. Extend the base resource controller—that's why it exists."
"Auth tokens go in httpOnly cookies, not localStorage. We had this discussion."
"Don't mock the database in tests. We use a real Postgres container—it's already in the docker-compose."
These corrections contain useful context—why I want something done a certain way. But they disappear when the session ends.
Superwiser watches your prompts (not Claude's responses, not tool calls, not file operations—just your prompts) and extracts preferences and corrections. Stores them locally. Claude can search them in future sessions.
How it's different from memory plugins: those capture everything Claude does and compress it into summaries. Works, but burns tokens and duplicates what Claude already handles. Superwiser only tracks human steering. Uses Sonnet in the background to extract rules, so it's cheap and fast.
Also has conflict detection. If you said "use Redux" three months ago but now say "just useState," it asks which one applies.
Simple rules still belong in CLAUDE.md. "Use TypeScript" doesn't need a plugin. Superwiser is for the contextual stuff that emerges during sessions—harder to anticipate upfront.
Install:
/plugin marketplace add mu4farooqi/superwiser
/plugin install superwiser
Just mention "use Superwiser" in your prompt to load relevant context. We don't pre-load silently in the background.
Experimental. I'm using it daily but expect rough edges. Contributions welcome.
