lythoskill

lythoskill

Declarative skill governance AND orchestration for AI agent teams. Declare which skills a project needs in skill-deck.toml. Combo prompts tell the agent how to orchestrate them. deck link reconciles the working set — undeclared skills are physically absent from the agent's view. The deck is the orchestrator: skill manifest + playbook + isolation discipline, all in one file. Built and self-governed by AI agents — see how we govern this project.

🌐 Languages: 中文

---

What problem does this solve?

When two skills both describe how to write tests, your agent gets conflicting instructions. "Disabling" a skill doesn't remove it — the agent can still see it. And when something goes wrong, you can't tell which skill caused it without removing them one by one. lythoskill enforces deny-by-default: undeclared skills are absent from the working set. If a conflict surfaces, bisect your skill-deck.toml — the lockfile tells you exactly what was loaded.

	npx skills add	lythoskill deck link
Install	npx skills add <repo> — imperative	Declare in skill-deck.toml — declarative
After install	Skill stays in scan path until remove	deck link reconciles — undeclared skills absent from working set
Multi-platform	-a claude-code / -a cursor per install	One TOML, configure working_set per platform
Discover	npx skills find → skills.sh directory	Curator scans local cold pool; agent uses WebSearch for discovery
Share	skills-lock.json	skill-deck.lock

Skills	State	Action
0–3, no conflicts	Simple	deck works, but manual management is fine too.
5–10, some conflicts	Growing	Use deck — declare which skills this project needs.
10+, you author skills	Ecosystem	Use deck + creator — thin-skill pattern.
Managing across teams/projects	Ecosystem	Full lythoskill: deck + creator + curator + arena.

skill-deck.lock gives you a single source of truth for "what skills were active when this was built." Commit it. Teammates get identical working sets. CI is reproducible. When something breaks, bisect the toml — the lockfile tells you exactly what changed.

---

🤖 For Agents

If you are an AI agent reading this README:

1. Check bun --version — if missing, install Bun
2. Run bunx @lythos/skill-deck@latest link --deck <url|path> to activate the project's deck
3. Skills load from the working_set directory in skill-deck.toml

Agents use lythoskill to build real things. showcase/ contains pages, reports, and tools built by agents using lythoskill-governed decks.

---

Quick Start

Give this to your agent

Copy the block below. Tell your agent: "Set up lythoskill for this project." The agent reads, installs Bun if needed, creates a deck, runs deck link, and verifies.

Read https://raw.githubusercontent.com/lythos-labs/lythoskill/main/README.md.
1. Check `bun --version` — if missing, install Bun.
2. Pick a skill from skills.sh or anthropics/skills.
3. Create `skill-deck.toml`:
   ```toml
   [deck]
   max_cards = 10
   [tool.skills.frontend-design]
   path = "github.com/anthropics/skills/skills/frontend-design"
   ```
   Then run `bunx @lythos/skill-deck@latest link`.
4. Verify skills appeared in your working_set directory.

That's the agent-era quick start: tell your agent, not your terminal.

Or run it yourself

curl -fsSL https://raw.githubusercontent.com/lythos-labs/lythoskill/main/examples/quick-init.sh | bash

The script installs Bun, creates a skill-deck.toml with frontend-design (Anthropic's official design skill), runs deck link, and self-checks. One command, a skill appears, done.

It's the same feeling as npm install — declared manifest, one command, reproducible result on disk. The difference from npx skills add: deck link is declarative (like package.json), not imperative (like brew install).

To add more: curl ... | bash -s -- --skill github.com/owner/repo/path. Or edit skill-deck.toml and re-run bunx @lythos/skill-deck@latest link.