Portable AI Agent Skills

Give agents a safer project evidence workflow.

Install inspectable instructions for recovery, Project Brain, universal inspection, change impact, evidence imports, release readiness, and extensions—without silently granting an agent new powers.

The bundle

Nine focused workflows, not one giant prompt.

Each skill is independently inspectable and can be installed only where it is needed.

restore-rehearsal

Validate and execute disposable Compose restore and candidate-upgrade drills.

project-brain

Build and query focused local project evidence without executing the project.

rehearse-init

Initialize an existing project without overwriting its work or inventing commands.

universal-inspector

Inspect web, native, CLI, API, code, documentation, and data surfaces with evidence.

change-impact

Compare structural snapshots and choose verification based on affected connections.

evidence-import

Normalize external test, security, contract, and infrastructure evidence with provenance.

recovery-adapters

Detect local exports and prepare offline recovery proposals for review.

release-readiness

Check packages, compatibility, SBOMs, checksums, and bounded release gates.

plugin-authoring

Build typed, bounded, explicit-trust importers, analyzers, and recovery extensions.

Choose the project directory

Install one skill or all nine.

The installer validates bundled files, refuses overwrites, and rolls back an incomplete multi-skill installation.

# Codex and agents using the common .agents directory
rehearse skill install .agents/skills --all

# Claude Code
rehearse skill install .claude/skills --all

# GitHub Copilot
rehearse skill install .github/skills --all

rehearse skill list --json
rehearse skill doctor --json

To install just Project Brain, replace --all with --name project-brain.

Capability boundary

Instructions do not equal permission.

Installing a skill supplies a workflow. It does not grant browser, desktop, Docker, network, deployment, secret, or repository-write access. The host agent and user remain responsible for authorization.

Outbound data

Keep project evidence private.

API keys, credentials, tokens, private keys, and backup contents must never be uploaded. Any other project information requires explicit consent for its destination and scope before it leaves the local project.

Review skill and plugin supply-chain guidance →

Local tool interface

Query a brain over local stdio MCP.

Rehearse also exposes a bounded local MCP server for an existing brain. It uses standard input and output rather than a hosted service.

rehearse mcp serve project.rrbrain
Review the host configuration: the MCP process remains local, but the connected agent’s own model, tools, and data-handling policy determine where returned context may go.