Steward

Steward is a portable Agent Skill that helps a coding agent discover what could make the current product dramatically more valuable, not merely what task is easiest to finish next.

It combines repository evidence with a compact vision council, separates operational obligations from ambitious bets, proposes falsifiable objective packets, and turns completed experiments into durable decisions and reusable leverage.

What it does

• Finds user friction, reusable assets, open product frontiers, and operational obligations.
• Generates a Big Swing, Smart Wedge, and compounding Flywheel before converging.
• Keeps maintenance visible without letting it consume the product-bet slot.
• Requires a visible objective packet and later explicit acceptance before making changes.
• Runs accepted work until its criteria are proven or a declared stop condition occurs.
• Closes experiments as Proven, Disproven, or Inconclusive and proposes an exact artifact promotion manifest.

Steward ships as plain instructions. It includes no hooks, MCP servers, credentials, pre-approved tools, or background processes. Repository instructions and host permissions remain authoritative.

Install for Claude Code

Exact /steward command

Install the Agent Skill at user scope:

npx skills@latest add manyrocks/steward --skill steward --agent claude-code --global --copy

Start Claude Code inside any repository and invoke:

/steward

Remove --global to install into the current repository at .claude/skills/steward instead. Review the files before committing a project-scoped installation.

On Windows, --copy avoids requiring symbolic-link privileges. To disable the installer's anonymous telemetry for the current PowerShell session:

$env:DISABLE_TELEMETRY = "1"

Managed Claude plugin

Add the marketplace and install the plugin:

claude plugin marketplace add manyrocks/steward
claude plugin install steward@manyrocks-skills --scope user

Reload plugins if Claude requests it:

/reload-plugins

Plugin skills are namespaced, so invoke:

/steward:steward

Use --scope local instead of --scope user to enable the plugin only for the current repository.

Install for Codex

npx skills@latest add manyrocks/steward --skill steward --agent codex --global --copy

Then ask Codex to use $steward, or describe a matching repository strategy request and let the skill trigger automatically.

Update

Agent Skills installation:

npx skills@latest update steward --global --yes

Claude plugin installation:

claude plugin marketplace update manyrocks-skills
claude plugin update steward@manyrocks-skills --scope user

Use

Common invocations:

/steward
/steward goal
/steward operate
/steward janitor
/steward close
/steward Here is the product vision I want to reconcile...

The default invocation is read-only and ends by asking you to choose among repository-specific futures. Selecting a future renders a complete objective packet. Work begins only after you accept that visible packet in a later turn.

Optional repository context

Steward derives product language and boundaries from the repository. For more deliberate guidance, add a repository-owned STEWARD.md containing only what the project wants every Steward run to know:

# Steward context

## Product promise

## Primary users

## Domain vocabulary

## Non-negotiable boundaries

## Useful local workflows

Repository instructions such as AGENTS.md and CLAUDE.md take precedence over this overlay.

Develop and validate

Validate the Agent Skill:

npx skills@latest add . --list

Validate and load the Claude plugin:

claude plugin validate .
claude --plugin-dir .

Inside the development session, invoke /steward:steward.

License

MIT