Agent Kit

Give this to your agent · Install · Profiles · Top Skills · CC Switch · Engineering guide

Give this to your agent

[!TIP] Paste this single sentence into Claude Code, Codex, Qwen Code, OpenCode, Pi, OpenClaw, or another shell-capable coding agent:

Open https://github.com/cubxxw/agent-kit/blob/main/docs/bootstrap.md and follow it end to end to safely initialize or fast-forward upgrade this machine for the current agent; preserve existing configuration and secrets, preview every change, run the verification gates, and report conflicts instead of forcing them.

That is the whole handoff. The protocol tells the agent how to detect its host, pick a profile, protect existing state, install, verify, and report evidence.

Why agent-kit

Agent setup usually drifts in three places: copied skills diverge, private runtime config leaks into dotfiles, and a bootstrap script silently replaces something important. Agent Kit gives those concerns explicit boundaries:

• One canonical skill tree. Every supported host sees the same reviewed files through symlinks.
• Safe repetition. Install and upgrade are idempotent, dry-runnable, and refuse copied directories or foreign links.
• Curated defaults, broad radar. Popularity helps discovery; license, current value, non-overlap, and executable review decide installation.
• Public/private separation. Skills, safe instructions, hooks, source pins, and templates can be public. Keys, auth, models, sessions, and machine state stay local.
• Verification before trust. Catalog validation, tests, public-boundary scanning, source pins, and host status are part of “done.”

Install

Native installer

Best for a machine you control. core means Claude Code + Codex; select one host by name when the agent should configure only itself.

git clone https://github.com/cubxxw/agent-kit.git "$HOME/.agent-kit"
cd "$HOME/.agent-kit"

./bin/agent-kit doctor --strict
./bin/agent-kit install --profile full-stack --tool core --dry-run
./bin/agent-kit install --profile full-stack --tool core
./bin/agent-kit status --profile full-stack --tool core

Supported native targets:

Use --tool all only when you intentionally want views for all six hosts.

Open skills ecosystem

The open skills CLI reaches 70+ agent integrations and discovers every accepted Agent Kit skill under skills/.

# Browse before installing
npm exec --yes [email protected] -- skills add cubxxw/agent-kit --list

# Example: install every accepted skill for one host
npm exec --yes [email protected] -- \
  skills add cubxxw/agent-kit --global --agent qwen-code --skill '*' --yes

Replace qwen-code with claude-code, codex, opencode, pi, openclaw, or another supported agent identifier.

Profiles

catalog.json is the source of truth. Every third-party entry records its repository, upstream directory, full commit SHA, tree SHA, and license.

The design stack has three deliberately separate jobs:

1. deepen-design branches product truth, narrative, architecture, and composition before implementation.
2. ui-ux-pro-max supplies searchable UI/UX and stack-specific evidence.
3. design-taste-frontend rejects common frontend clichés at final preflight.

Taste rules alone can remove obvious slop while still converging on a polished template. Agent Kit therefore never treats a linter, Lighthouse score, or self-authored design score as proof of distinctiveness.

Give a design Agent this sentence when the first result is merely polished:

Use $deepen-design to audit the rendered interface, branch from product truth into two materially different directions, compare before/A/B evidence with the logo hidden, backtrack when both branches remain generic, and only after selecting an ownable architecture use $ui-ux-pro-max and $design-taste-frontend to implement and preflight it.

Top, without the bloat

docs/top-skills.md organizes high-signal GitHub skill sources into adopted, on-demand, and discovery-only layers. It covers official OpenAI, Anthropic, Vercel, Microsoft, Hugging Face, NVIDIA, .NET, Supabase, Firebase, Prisma, and Remotion sources plus strong specialist and community collections.

The radar can be broad because it installs nothing. The top profile remains small because every included skill must pass license, source-pin, executable, overlap, public-boundary, and host-discovery gates.

./bin/agent-kit install --profile top --tool core --dry-run
./bin/agent-kit install --profile top --tool core

How one source reaches every agent

flowchart LR
    G["GitHub: cubxxw/agent-kit"] --> C["Local canonical checkout"]
    C --> K["catalog.json profiles"]
    K --> S["skills/<name>"]
    S --> A["Claude Code"]
    S --> B["Codex"]
    S --> Q["Qwen Code"]
    S --> O["OpenCode"]
    S --> P["Pi"]
    S --> W["OpenClaw"]

The host directories are discovery views, not storage. Editing or fast-forwarding the canonical checkout updates every managed link.

Agent Kit + CC Switch

These projects solve different layers and work well together:

In CC Switch, open Skills → Repository Management → Add Repository, then use:

Owner: cubxxw
Name: agent-kit
Branch: main
Subdirectory: skills

For one shared source, select ~/.agents/skills as the CC Switch skill storage location and use symlink distribution. Keep provider credentials and CC Switch cloud-sync data out of this repository.

What is shared—and what never is

Files under config/ are mergeable examples. They are never a license to replace an existing settings.json, config.toml, or agent instruction file.

The engineering practices behind it

Agent Kit turns recurring corrections into infrastructure:

1. Keep resident context small; load detailed guidance only when it is needed.
2. Encode repeated mistakes as a rule, test, hook, skill, or script.
3. Make every completion claim return with observable evidence.
4. Preview first, preserve conflicts, and keep changes reversible.
5. Automate the deterministic path; reserve model judgment for real decisions.
6. Review third-party skill code, license, overlap, and source pin before use.

The dated research and trade-offs are documented in docs/best-practices.md, drawing from Boris Cherny’s workflow thread, the evolving How Boris Uses Claude Code, CC Switch, UI UX Pro Max, and Addy Osmani’s Agent Skills, with Taste Skill as the reviewed frontend judgment layer.

Operate it

# Safe preview
./bin/agent-kit install --profile full-stack --tool core --dry-run

# Validate repository, catalog, licenses, and public boundary
./bin/agent-kit doctor --strict

# Check managed links
./bin/agent-kit status --profile full-stack --tool core

# Check whether a pinned upstream skill directory changed
./scripts/check_upstreams.py

# Remove only links owned by this checkout
./bin/agent-kit uninstall --profile full-stack --tool core --dry-run

For unattended provisioning, set AGENT_KIT_PROFILE and AGENT_KIT_TOOL, then run scripts/bootstrap.sh. It only fast-forwards a clean checkout.

Trust, maintenance, and quality

• Architecture
• Bootstrap protocol
• Maintenance runbook
• Skill selection ledger
• Top Skills Radar
• Boris-style quality review
• Security policy
• Third-party notices

If this saves you from maintaining the same agent setup six times, consider starring the repository. It makes the project easier to find without making the catalog any less selective.

License

First-party code and documentation are MIT licensed. Vendored skills retain their upstream licenses and attribution; see THIRD_PARTY_NOTICES.md.