SKILL.md

---
name: build-an-agent
description: Use when the user wants to build, configure, or publish an Outside Agent — a hosted AI agent that talks to end users over SMS, voice, and web. Drives the outside-agent MCP tools end to end.
---

# Build an Outside Agent

You have the `outside-agent` MCP server connected. It lets you compose a hosted consumer
AI agent for the signed-in creator and publish it to SMS, voice, and web.

## Interview first, don't build blind

A short brief always leaves design decisions unmade. Before publishing, walk the design
tree WITH the creator:

- The build tools (`createAgent`, `createConversationalAgent`, `setAuthoredConfig`,
  `patchAuthoredConfig`) return a `design.openDecisions` array — the dimensions still
  unpinned (purpose, audience, soul, capabilities, guidance, channels), each with the
  question to ask AND a recommended answer.
- Put those questions to the creator ONE AT A TIME, in order, each with your recommended
  answer, and wait for their reply before the next.
- If a decision is answerable by exploring (`getCatalogSummary` for what it can do,
  `listConnectors` for an integration), explore and decide instead of asking.

## Workflow

1. `createAgent` (name it) — or `createConversationalAgent` for a plain persona agent in one call.
2. `recordSourceRequirements` — capture the raw brief + each interview Q&A before you review.
3. `setAuthoredConfig` / `patchAuthoredConfig` — write the whole config (persona, soul,
   capabilities, guidance, channels).
4. `seedKnowledgeText` — ground the agent in the creator's real content if it answers questions.
5. `lintConfig` → `simulateBuild` → `reviewAgent` — validate and safety-review.
6. `publishAgent` (or `publishAndDeploy`) — ship it. Confirm any `assumedDefaults` it returns
   with the creator rather than treating publish as the end.

Prefer `getCatalog` / `getCatalogSummary` to learn the real capability + connector catalog
before configuring, so you author against what the platform actually supports.