Guides users through creating a new agent via conversational interview
Agent Creator guides you through designing a custom Metaflow agent via a structured conversational interview. Instead of dumping instructions into a blank system prompt, you define the agent's purpose, tone, tools, knowledge boundaries, escalation rules, and success criteria step by step. The skill translates your answers into a production-ready agent profile aligned with how Metaflow runs tasks, calls tools, and applies other skills.
It covers persona definition, workflow triggers, output formats, guardrails for sensitive topics, and when to delegate to specialized skills like SEO Audit or Google Ads Optimizer. The result is an agent teammates can reuse for recurring work—weekly reporting, campaign QA, client onboarding—not a one-off chat configuration.
Marketing ops leads standardizing how teams use AI, agency owners productizing delivery with consistent agent behavior, and practitioners who know their process but want help encoding it. Useful when off-the-shelf assistants hallucinate process steps or ignore your company's naming conventions and approval chains.
Also fits teams onboarding new hires: the agent becomes a documented, executable version of senior practitioner judgment for common scenarios.
The interview progresses through discovery (what problem does this agent solve?), scope (what it should never do), inputs (data sources, files, integrations), outputs (Slack summaries, Sanity drafts, HTML artifacts), and skill attachments. You refine edge cases: incomplete data, conflicting KPIs, brand voice for enterprise versus SMB audiences.
Review the draft agent charter: role statement, decision tree, example user prompts, and quality checklist. Test with three realistic tasks, note failures, and tighten instructions. Document maintenance—who updates the agent when offers or platforms change.
People ask how Agent Creator differs from writing a long custom GPT instruction. Here the output maps to Metaflow's agent model: skills, tools, account context, and team sharing. Another question is how specific to be; the skill balances concise rules with exemplar prompts so the agent generalizes without drifting.
Teams wonder if one agent can cover SEO and paid media. The workflow recommends primary specialization with explicit handoff patterns to other skills rather than one overloaded generalist.
Start a task with Agent Creator enabled and answer the interview prompts honestly—include real examples of good and bad outputs. When the draft is ready, save or deploy per your workspace workflow and run pilot tasks before rolling out to the team. Reopen Agent Creator when you add new integrations or skills to keep the agent current.
For broader context, see our roundup of claude skills marketing, and read how to create Claude skills for related setup guidance.