UnleashedPPOS/unleashed-superpowers

A Claude Code plugin that catches human-team metaphors in AI-system design and forces AI-native reframing. Plus a reference-quality feature-documentation skill that enforces source-grounded accuracy.
The thesis
Most people designing AI systems accidentally design digitised human teams. They import agent roles, seniority framing, calendar estimates, and v1/v2 roadmaps — metaphors that silently delete better architectures from the design space.
This plugin catches that and reframes around AI-native constraints (tokens, latency, observability, determinism, parallelism). The four-principle summary:
1. **Don't anthropomorphize the structure.** No team-of-agents architecture. Use AI-native patterns (blackboard, DAG, ensemble, queue, sub-agent spawn).
2. **DO absorb human expertise as content.** Mine what humans know — especially their mistakes — and encode it as critic rules, validation schemas, few-shot examples.
3. **Documentation is executable.** SOPs, runbooks, and lessons-learned files are the persistent form of human expertise. The AI reads them, is tested against them, and proposes updates.
4. **Parallelism is the default.** Sequential is the exception that must justify itself.
Design thesis: *Anthropomorphic surface, AI-native engine, event-log bridge.*
Full argument: docs/specs/2026-04-18-two-layer-ai-native-design.md.
Installation
Option 1 — Claude Code plugin install (preferred)
After install, the skills auto-register in every new Claude Code session.
Option 2 — Git clone + symlink (fallback if Option 1 is not yet supported)
Start a fresh Claude Code session. The skills appear in the registry.
What's inside
Skills
**designing-ai-systems** — Four-pass AI-native reframe.
Triggers automatically on phrases like "agentic system," "multi-agent," "50 researchers," "team of agents," "senior reviewer agent," or any design discussion for LLM-coordinated systems. Runs four passes in order:
- **Pass 0** — mine human domain expertise (mistakes, edge cases, hard rules, vocabulary)
- **Pass 0.5** — design the documentation substrate (SOPs, runbooks, lessons-learned, glossaries)
- **Pass 1** — engine reframe (AI-native patterns, parallelism-as-default, stress-test failure modes)
- **Pass 2** — presentation design (anthropomorphic surface, event-log bridge, SOP viewer)
Mantra: *"Humans as content. Documentation as fuel. Parallelism as default. Anthropomorphic surface, AI-native engine, event-log bridge."*
**document-feature-module** — Reference-quality feature documentation.
Produces docs a new contributor (or AI agent) can trust without cross-checking the source. Two lenses: Senior Engineer (every code claim must grep to a real symbol; invariants justified with Why: lines; failure modes and perf characteristics named) and Senior Designer (scanability, verb-first copy, clear extension points).
Golden rule: **every factual claim resolves to a grepable symbol.** If you can't grep it, don't write it.
**high-level** — High-Level mode: end-state-only reporting.
Like caveman mode, but you keep normal sentences and drop everything the reader doesn't need. No narration while the agent works, then one final message: **Done / Remaining / Heads-up / Proof / Needs you**, one line per item, each fact said once, with no line cap so open items are never dropped. Say "high-level mode" (or /high-level) to turn it on and "normal mode" to turn it off.
To make it permanent, run **/high-level-mode on** once. That writes a flag file (~/.claude/high-level-mode), and the plugin's SessionStart hook (hooks/high-level-session-start.sh) then loads the rules into every new session: terminal, desktop app and IDE, plus after /clear and compaction (resumed chats already have them). /high-level-mode off turns it off; HIGH_LEVEL_MODE=0 skips it for one session. Use the hook or the output style, not both, or the rules load twice. This is the recommended route because the desktop app fixes a session's output style when the session starts and only lists custom styles in Settings.