Skip to content

Agent-first workflow

Three skills cover the agent surface: growthos-sprint-dev (tickets), growthos-product-context (seed living PRD from a repo), growthos-prototype-design (UX Spec → Figma MCP, write figmaUrl back). Pair Figma’s remote MCP with the last one.

The sprint skill is one task per turn: ask before claim, implement, ## Agent handoff, DONE, stop. Ask before advance_sprint. Humans still own PRD finalize, UX Spec approve/skip, sprint start, and Connected policy.

  • Who: Engineer or founder who wants the agent to drive tickets
  • Time: After keys + an ACTIVE sprint (or living PRD explore)
  • Plan: Free: 1 MCP handoff. Pro+: unlimited handoffs. Sprints still Pro+.
  • Keep one gos_ key per human, never a shared company key.
  • Switch skills: explore repo, design in Figma, then sprint loop.
  • Stop the agent for judgment calls (scope, skip UX Spec, Connected).
  • You trust the board more than a chat transcript as the work queue.
  • You want MCP as the default interface after planning.
  • MCP connected at https://mcp.growthos.salesshuru.com/mcp.
  • For sprint loop: ACTIVE sprint and claimable Native tasks.
  • Settings → API Keys & Agents — config, prompts, SKILL.md downloads.
  • Sprint board → Connect coding agent.
  1. Mint a key and paste Cursor/Claude config. Confirm MCP auth.
  2. Download the skill you need this hour (sprint, product-context, or Figma).
  3. State the project id. Let the agent propose; you confirm.
  4. One task (or one seed / one Figma screen batch as the skill says). Then stop.
  5. Humans: finalize PRD, approve phases, start sprint, decide Connected vs Native.
  • Each agent turn leaves an audit trail on the ticket or living PRD.
  • You did not bulk-claim the sprint.
  • On Connected projects, never overwrite synced status.
  • Product-context is for next-wave living PRD, not a rewrite backlog.
  • Running all three skills in one vague prompt.
  • Skipping confirm-before-claim because you are “agent-first”.