Skills and agents for AI coding assistants — Claude Code, Codex, Cursor, OpenCode, Copilot, Gemini CLI and more. One install, every harness.
A skill is a folder with a SKILL.md that teaches an agent how to do one job well. An agent is a subagent definition with its own model, tools, and conduct. This repo holds the ones I use every day: verify visual output before calling it done, publish a page that survives being shared, turn arXiv papers into a decision, route each task to the cheapest model that can do it, build decks and themes that look like one system.
Pick one. All three land in the same place and can be re-run safely.
# 1 · The standard way (works for ~20 harnesses, interactive picker)
npx skills add vimoxshah/skills
# 2 · This repo's installer — detects your harnesses, asks what to link, symlinks from one central clone
npx @vimoxshah/skills
# 3 · No Node? Same thing in bash
curl -fsSL https://raw.githubusercontent.com/vimoxshah/skills/main/install.sh | bash
Claude Code users can also install it as a plugin:
/plugin marketplace add vimoxshah/skills
/plugin install skills@vimoxshah
Want just one skill? npx skills add vimoxshah/skills --skill visual-verify
Non-interactive (CI, dotfiles): npx @vimoxshah/skills --yes --all
Running from a clone of this repo? Use
node bin/install.mjs(ornpm start) instead. Inside the checkout,npx @vimoxshah/skillsresolves to the local copy and fails withcommand not found: vimoxshah-skills, because the bin is only linked after annpm install. Thenpxform is for using the package, not for working on it.
| Skill | What it does |
|---|---|
visual-verify |
Render the page, chart, diagram, or deck and look at it before reporting it works. Covers the animation-settle trap and three verification tiers. |
html-artifact |
Build single-file HTML pages that survive being shared: no network, light + dark that honours the viewer's toggle, deck shells, SVG diagrams, a pre-publish check. |
arxiv |
Turn arXiv research into a decision for the repo you are in — sweep a category, deep-read one paper's LaTeX, or map a corpus onto the codebase. Citation-gated: no figure is ever paraphrased into existence. |
claude-router |
Route each task to the right Claude tier — Fable for judgment, Opus to orchestrate and review, Sonnet to build, Haiku for volume — via model-pinned subagents. |
codex-orchestrator |
Claude plans and verifies; Codex executes. Bounded work packets, three Codex profiles, an escalation path. |
design-kit |
One token registry for every surface: 11 themes × 9 font pairings validated as a pair, a font pipeline with a license manifest, motion tokens, chart ramps. Compiles to decks, HTML artifacts, video frames, and pptx. |
house-decks |
A self-contained 1920×1080 HTML deck system: keyboard nav, presenter window with speaker notes, print-to-PDF variant, reveal/pop/terminal animations, and a narrative guide for picking the arc. |
explain-interface |
"How was this built?" — give it a URL and the thing you're curious about; it reads the layers that produce the effect and explains each one. |
scoped-worktree-delivery |
Git worktrees done safely: prove the topology, verify author identity, isolate the change, publish only when authorized. |
eli5 |
/eli5 <topic> — a dead-simple picture explainer as an HTML artifact. |
Model-pinned lanes the router dispatches, plus specialist advisors.
| Agent | Model | Role |
|---|---|---|
advisor |
Fable | Read-only judgment at commitment boundaries — architecture, migrations, contracts. |
reviewer |
Opus | Read-only diff review and synthesis; hunts weakened tests and planted shortcuts. |
hard-implementer |
Opus | The write lane for hard-as-code changes and root-cause debugging. |
implementer |
Sonnet | The everyday build lane; writes code and proves it with tests. |
explorer |
Haiku | Fast read-only fan-out: locate, map, summarise. |
paper-analyst |
Opus | Deep-reads one arXiv paper's LaTeX and returns a typed relevance verdict. |
presentation-director |
Opus | Takes "I need to present X to Y in N minutes" and delivers the finished piece. |
multi-agent-architect |
Sonnet | Topology, trust, failure recovery, and HITL gating for multi-agent systems. |
statistician · technical-writer · ux-architect · brand-guardian · content-creator · developer-advocate · database-optimizer · identity-access-engineer · search-relevance-engineer · persona-walkthrough |
Sonnet | Specialist advisors. Adapted from agency-agents (MIT). |
/style picks an output style from a popup; two styles ship (attention-kind — ADHD-friendly, front-loaded; plain — answer first, warm, short); a statusline script showing model, effort, context used, and cost; the two routing policies (CLAUDE-ROUTING.md, CLAUDE-CODEX-ROUTING.md).
rules/context7.md is a standing rule that sends library and API questions to live documentation instead of training data. The installer does not edit your CLAUDE.md — add it yourself with @rules/context7.md if you want it.
| Harness | Skills | Agents |
|---|---|---|
| Claude Code | ~/.claude/skills |
~/.claude/agents (native) |
| OpenAI Codex CLI | ~/.agents/skills |
~/.codex/agents/*.toml (transpiled) |
| Cursor | ~/.cursor/skills |
~/.cursor/agents (transpiled) |
| OpenCode | ~/.config/opencode/skills |
~/.config/opencode/agent (transpiled) |
| GitHub Copilot CLI | ~/.copilot/skills |
installed as skills |
| Gemini CLI | ~/.gemini/skills |
installed as skills |
| Goose · Amp · Factory · Kiro · Windsurf · Cline · Roo | their skills dir | installed as skills |
Paths come from harnesses.json. If your harness is missing, add a row there and open a PR.
~/.agents/vimoxshah-skills/ ← one shallow clone (the central place)
├── skills/visual-verify/
└── agents/reviewer.md
~/.claude/skills/visual-verify → symlink into the clone
~/.cursor/skills/visual-verify → symlink into the clone
~/.codex/agents/reviewer.toml ← copy, transpiled to Codex's format
- One copy on disk. Every harness reads the same files.
updateis agit pull; the links stay valid. - Never clobbers. A real directory in the way is skipped with a warning, not replaced. Your edited routing policy is never overwritten.
- Reversible. A receipt records every path created;
uninstallremoves exactly those. - Windows defaults to copy (symlinks need Developer Mode).
npx @vimoxshah/skills update # pull the latest
npx @vimoxshah/skills doctor # broken links, drift, unwired harnesses
npx @vimoxshah/skills uninstall # remove everything the installer created
npx @vimoxshah/skills list # what's in each bundleBundles: core · routing · research · design · experts · claude-extras. Defined in manifest.json.
Each skill follows the Agent Skills spec — a directory with a SKILL.md whose frontmatter carries name and description, plus any scripts or references the skill loads on demand. That is why npx skills add works with no code from this repo.
skills/visual-verify/
├── SKILL.md # name, description, the instructions
└── scripts/ # optional helpers the skill calls
Issues and PRs welcome. Run node scripts/lint.mjs before pushing — it checks frontmatter, bundle membership, and that nothing machine-specific leaked in. Keep a skill to one job; if it needs a second SKILL.md, it is two skills.
MIT © Vimox Shah. Third-party notices in THIRD_PARTY.md.