Bot Generator Bot turns 39 old prompts into typed, installable skills

What it is

You describe a bot idea, pick a template, define the typed output you want, and get a runnable MetaHarness agent — exported standalone or driven through the CLI or MCP. Compilation stays offline; model calls need explicit operator configuration.

What changed

Pull request #9 completes the migration: all 39 original prompts now have a curated typed template, an installable skill and a dedicated MCP builder. The catalog maps each original source to its skill and tool, and the original text is untouched so provenance is visible.

The tool surface is 50 tools, 42 prompt templates and three resources.

Get started

Node.js. List the skills the repository publishes:

npx -y skills@1.5.25 add ruvnet/Bot-Generator-Bot --list

Install one into both Claude Code and Codex, project scoped:

npx -y skills@1.5.25 add ruvnet/Bot-Generator-Bot --skill bot-generator \
  --agent claude-code codex --copy --yes

Expected result: .claude/skills/bot-generator for Claude Code and .agents/skills/bot-generator for Codex. Invoke /bot-generator in Claude Code or $bot-generator in Codex.

Register the MCP server separately — a skill alone gives guidance, not tools:

git clone https://github.com/ruvnet/Bot-Generator-Bot.git
npm install --prefix Bot-Generator-Bot
claude mcp add --transport stdio --scope project bot-generator \
  -- node /absolute/path/Bot-Generator-Bot/src/cli.mjs mcp

Pinning the installer version does not pin the repository revision. If reproducibility matters, keep the reviewed checkout commit.

Use it today

Practical case: a repeatable internal assistant. Input is a template id and your typed output contract; the workflow is compile offline, review the manifest, then wire a model; the output is a validated agent manifest with a stable SHA256.

Acceptance test: compile the same specification twice. The manifest hash must be identical. If it is not, something non deterministic slipped into your spec.

Push it further

Experimental commentary. A catalog of typed agents is the raw material for an agent that picks agents. The limitation is stated plainly in the repository: these skills draft, assist and simulate. They do not deploy, transact or provide licensed professional advice.

Falsifiable test: ask a generated agent for an action outside its declared tool policy. The request should be refused by the policy, not by the model's manners.