x-skills
Intelligent skill routers for Claude Code. Each skill classifies user intent and routes to the optimal executor — with optional multi-model orchestration via OpenCode.
All dependencies are optional. Skills degrade gracefully to Claude-only mode when external tools aren't available.
Install
/plugin marketplace add quangtran88/x-skills
/plugin install x-skills@x-skills-marketplace
/reload-plugins
Then run setup to configure the bindings (omo-agent, agy-agent) and detect available dependencies:
/x-skills:setup
Upgrading
After every plugin upgrade, re-run /x-skills:setup. It refreshes:
- Symlinks (
~/.local/bin/omo-agent, ~/.local/bin/agy-agent) to point at the new plugin cache
~/.config/x-skills/capabilities.json (new fields appear with each release)- Detection of newly-supported dependencies (e.g. v1.4.0 added
gemini CLI detection for the x-gemini skill)
The plugin's SessionStart hook detects stale capabilities and reminds you, but you can run /x-skills:setup proactively any time — it's idempotent.
Skills
| Skill | Invoke | What it does |
|---|
| x-do | /x-skills:x-do | Universal execution router — classifies tasks into modes (new feature, bugfix, quick edit, refactor, plan execution) and dispatches to the right workflow |
| x-research | /x-skills:x-research | Universal research router — classifies questions and routes to optimal agents (codebase search, external docs, architecture review, OSS internals, pre-planning) |
| x-review | /x-skills:x-review | Code/plan/PR review orchestrator — cross-model review with Claude + GPT perspectives, structured verdicts |
| x-verify | /x-skills:x-verify | Completion-cascade dispatcher — answers "am I done?" for long-running skills with mandatory fallback to prevent silent success claims |
| x-bugfix | /x-skills:x-bugfix | Structured debugging — routes through investigation, hypothesis testing, and verified fix with evidence collection |
| x-mindful | /x-skills:x-mindful | Pre-implementation architect-review gate — walks the user through the tradeoffs AI silently picked, assumptions it baked in, blind spots, over/under-engineering, and future-debt at architect level (never code-level), one item at a time |
| x-design | /x-skills:x-design | Visual design system integration — resolves brand references (Linear-like, Stripe-like) to curated DESIGN.md files from 58 indexed sites |
| x-api-pentest | /x-skills:x-api-pentest | API security testing — OWASP API Top 10 testing with schemathesis, nuclei, sqlmap, spectral |
| x-qa | /x-skills:x-qa | Profile-driven E2E QA — scans the project once, persists a launch/test profile, fans tests out across cheap/strong models against an isolated stack |
| x-team | /x-skills:x-team | Multi-feature team orchestrator — decomposes work into N parallel features, provisions a worktree per feature, gates merges on x-qa passing |
| x-worktree | /x-skills:x-worktree | Isolated git worktree provisioner — used by x-do/x-bugfix --wt flag and directly; auto-applies docker-compose isolation when a profile is present |
| x-worktree-isolate | /x-skills:x-worktree-isolate | Per-worktree docker-compose isolation — scan once, emit profile.json, write compose.override.yml + .env.worktree per new worktree |
| x-omo | /x-skills:x-omo | OpenCode multi-model bridge — dispatch to GPT-5.5, Gemini, Codex models via role agents or direct model routing |
| x-gemini | /x-skills:x-gemini | Direct Gemini CLI bridge — uses Google Ultra subscription (no API key), native Google Search grounding, gemini-3.x access without OpenCode layer |
| x-guide | /x-skills:x-guide | Progressive comprehension-gated tutor — turns docs/PRDs/specs/code/features into a resumable walkthrough with per-project state in .x-guide/<slug>/ |
| x-skill-improve | /x-skills:x-skill-improve | Session-based skill alignment analyzer — evaluates how well a skill was followed during real sessions and proposes targeted edits |
x-shared is a reference library used by other skills (not invokable directly).
How It Works
Each skill is a router, not an executor. It:
- Reads
~/.config/x-skills/capabilities.json to detect available tools
- Classifies the user's request into a type/mode
- Dispatches to the best available executor (OMO agent, OMC agent, native Claude agent, or direct tool)
User: "add auth to the API"
│
x-do classifies → Mode A (new feature)
│
┌────┴────────────────────┐
│ opencode available? │
├─── yes ─→ OMO oracle │ (oracle review)
│ + OMC executor │ (Claude implementation)
├─── no ──→ Agent(opus) │ (Claude-only fallback)
└─────────────────────────┘
Dependencies
