libra-skills

The context layer for your agents.

Turns AI chat decisions into structured repo docs so Cursor and Claude Code always start with the right context. init-libra scaffolds the full structure; update-libra keeps it in sync.

AGENTS.md
ARCHITECTURE.md

docs/
├── decisions/
│   ├── INDEX.md
│   └── adr-NNN-slug.md
├── specs/
│   ├── INDEX.md
│   └── <feature-slug>.md
├── design/
│   ├── INDEX.md
│   └── core-beliefs.md
└── plans/
    ├── INDEX.md
    ├── active/
    │   └── <plan-slug>.md
    └── completed/
        └── <plan-slug>.md

Every doc includes YAML frontmatter. Cross-links use wiki-style notation: [[adr-005]], [[spec:auth-flow]]. Index files are always INDEX.md.

Installation

Claude Code

Add the marketplace to Claude Code (run inside an active claude CLI session):

/plugin marketplace add libra-mcp/libra-skills

Install the plugin:

/plugin install libra-skills@libra-mcp

Restart Claude Code, then run:

/libra-skills:init-libra

Cursor

Local development install from repo root (temporary until libra-skills is available in the Cursor plugin marketplace):

./scripts/install-local.sh

Restart Cursor (full quit/reopen), then run:

/init-libra

Usage

Claude Code commands (v1)

• /libra-skills:init-libra
• /libra-skills:update-libra

Cursor commands

• /init-libra
• update-libra is run in-stream when prompted by rules or when explicitly requested

Rules and hooks behavior

Claude Code (hybrid hooks + skills)

• SessionStart hook reminder: orient on Libra docs before planning/building
• TaskCompleted hook reminder/check: nudge docs sync when meaningful non-doc changes exist
• SessionEnd hook reminder/check: nudge docs sync before finishing

In Claude Code v1, hooks are lightweight nudges/checks and do not force doc writes.

Cursor (rules-first)

• rules/libra-docs.mdc: docs orientation (always-on)
• rules/update-libra.mdc: when to run docs sync (ask-triggered or session-end with meaningful changes)

License

MIT