Most "AI plugin marketplaces" ship browser extensions and prompt collections. Claude Code plugins are a different shape: each one ships procedural memory — the workflow, the API surface, the gotchas a senior engineer would teach you on day three — packaged so Claude loads it on demand and forgets it when irrelevant. 27 plugins · 58 skills · install in two commands.

/plugin marketplace add aiocean/claude-plugins
/plugin install <plugin-name>@aiocean-plugins

Skills don't activate until Claude sees a fuzzy match against your message, so installing things you might need later costs nothing at runtime.

Five jobs these plugins solve

Plugins cluster around real engineering work, not feature checklists:

• Translate & write — literary EPUB pipelines with cross-chapter glossary management, Vietnamese style enforcement (Tín-Đạt-Nhã), bilingual export. See aio-epub-translate, aio-epub-vn-style.
• Debug & investigate — race-condition surfacing, hypothesis-PoC loops, CDP-relay browser automation that doesn't fight your login state. See aio-cdp-relay, aio-ai-slop-cleaner.
• Design & ship UI — brutalist + neobrutalism systems with stamp shadows, WCAG audits, React-19 effect refactoring against the new useActionState / useOptimistic model. See aio-design-system, aio-dashboard-design.
• Operate at scale — StarRocks query tuning + EXPLAIN reading, Go production hardening (concurrency, generics, race detection), threat modeling for AI/ML workloads, multi-agent orchestration patterns. See aio-starrocks, aio-golang-mastery, aio-threat-models, aio-software-architect.
• Build for yourself — skill creation flow, CLAUDE.md tuning, marketplace publishing, prompt patching for Claude Code itself. See aio-claude-toolkit.

Browse the 27-plugin catalog → · Read the guides →

Why these primitives compose

A skill is procedural knowledge that loads when Claude sees a description match. An agent is an isolated sub-thread with its own context window. A hook fires on tool-call events (PreToolUse, PostToolUse, Stop). When you chain all three — skill informs Claude what to do, agent runs the work in isolation, hook validates the output — you stop writing "please follow this checklist" prompts and start shipping reproducible workflows.

The primitives guide shows the decision tree. The CLAUDE.md guide explains why project memory is not documentation. The install guide covers the rest.