Agent skills for OpenCode. Install with npx skills or clone directly.
| Skill | Description |
|---|---|
| design-exploration | Generates multiple design variants with different visual directions, then implements the chosen one |
| grok-cli | Use the Grok CLI for fresh, real-time info — deep X/Twitter digging and live news/web research, with a verified one-command recipe |
| opencode-memory | Read-only access to local OpenCode history — sessions, plans, conversations, and project context |
| x-algorithm | Write X posts that the For You algorithm rewards — grounded in the open-sourced ranker, retrieval, and content classifiers |
Generates multiple distinct design variants of a component or page, each with a completely different visual direction, then implements the chosen one.
npx skills add carson2222/skills --skill design-exploration -g -yOr browse on skills.sh.
Give it a component, page, or section and it produces N variants (default 5), each taking a fundamentally different visual direction. You compare them in the browser, pick one, and it implements the final version with real data.
Works in two modes:
- New project — no design system yet, variants define one. Each gets its own fonts, colors, layout philosophy.
- Existing app — design system already exists. Variants explore different layouts and compositions while staying consistent with your tokens and patterns.
Claude Opus 4.6 with high thinking budget covers ~95% of runs. Gemini can produce solid UI too. Open source models (GLM5, MiniMax, Kimi K2.5) sometimes surprise you but are less predictable. Codex is not recommended for frontend design.
This is not a "run it and pray" skill. Output quality scales directly with context. Before running:
- What matters to you? Professional? Playful? Dark and moody?
- Who is this for? What vibe are you going for?
- Drop assets (images, logos, icons) in the project folder with descriptive filenames.
Talk to it like you'd talk to a designer. The more you give it, the better it gets.
Use the Grok CLI as a sub-agent to pull fresh, real-time information the model doesn't have — especially digging deep into X/Twitter and breaking news.
npx skills add carson2222/skills --skill grok-cli -g -yOr browse on skills.sh.
When you tell the agent to "use grok", it burns several turns rediscovering whether grok exists, what flags it takes, and how to call it — then often hangs. The default model loops on web_fetch against X/news URLs and never returns, so the call dies on a timeout with nothing to show.
This skill hands the agent one verified command and tells it to run it immediately. No discovery dance, no hung calls.
Teaches the agent to call grok -p headless for live data, with the exact recipe that actually works:
-m grok-build— the Grok model, not the composer model--yolo— auto-approve tools (required headless)--disallowed-tools web_fetch— force xAI liveweb_search; the key fix that stops the fetch loop--max-turnscap + JSON output parsed withjq
Includes ready-to-paste templates for deep X/Twitter digging (specific posts, @handles, URLs, sentiment, influential accounts) and live news research, plus follow-up calls in the same grok session.
- The
grokCLI installed and authenticated (curl -fsSL https://x.ai/cli/install.sh | bash, thengrok login) jqon PATH (pre-installed on most macOS/Linux)
"use grok", "ask grok", "dig into twitter/x", "what's X saying about…", "latest/fresh on…", or autonomously whenever current real-time data beyond the training cutoff would help.
Lightweight, read-only access to your local OpenCode history. No injection, no bloat — the agent can look things up when it would help.
npx skills add carson2222/skills --skill opencode-memory -g -yOr browse on skills.sh.
Every "memory" plugin out there injects context everywhere, bloating every request. Most of the time you don't need history. But when you do — resuming a project, referencing a past decision, debugging something that came up before — the agent should be able to check on its own without you having to explain where OpenCode stores things.
Without this skill, getting the agent to browse local OpenCode data requires precise wording and a lot of luck. It tries Claude, Codex, random paths, and gives up. This skill solves that by teaching the agent exactly where to look and how to query it.
- Sessions — list, filter by project, search by title
- Messages — read full conversations from any session
- Search — full-text search across all past conversations
- Plans — list and read saved plans
- Projects — see all tracked projects with session counts
- Prompt history — recent prompts you've typed
The skill teaches the agent the OpenCode storage model and provides ready-to-use sqlite3 queries. It uses opencode db path when available to find the active database, then opens SQLite read-only. No scripts, no external tools — just sqlite3 (available on every Mac/Linux) and bash. The SKILL.md contains the database paths, schema, and copy-paste query templates for every common operation.
The agent will auto-trigger when:
- You reference something done before ("last time", "that plan we made")
- You're resuming work on a project
- You ask about history, plans, or past sessions
- It suspects prior context would help
sqlite3CLI (pre-installed on macOS and most Linux)- OpenCode must have been used at least once (creates the database)
- macOS/Linux (reads
~/.local/share/opencode/, respectsXDG_DATA_HOME)
Write X (Twitter) posts that the For You algorithm actually rewards. Derived directly from the open-sourced recommendation system — the Grok-based transformer ranker, Phoenix retrieval, Thunder in-network store, and Grox content-understanding pipeline.
npx skills add carson2222/skills --skill x-algorithm -g -yOr browse on skills.sh.
Teaches the agent the actual scoring mechanics of the X For You feed so it can write, review, and debug posts. Covers the 19 predicted engagement actions and their weights, hard filters that silently kill posts, out-of-network retrieval mechanics, author diversity decay, content quality classifiers (VLM banger screen, spam detection), and the engagement caching window.
Use it when you want to:
- Write a post optimized for a specific goal (reach, replies, follows, shares)
- Review or rewrite an existing draft with algorithm-aware feedback
- Diagnose why a post underperformed
- Plan a content strategy grounded in how distribution actually works
- Full breakdown of the scoring formula (positive + negative action weights)
- Pre-post checklist derived from the actual pipeline
- "Why did this flop?" diagnostic sequence
- Anti-patterns the algorithm punishes
- Myth-busting section for X folklore
Based on the X For You Feed Algorithm open source release (May 2026).
Apache 2.0. See LICENSE.