Skip to main content

Vibe coding

Vibe coding means describing the agent you want in everyday language and letting an AI assistant (Cursor, Claude Desktop, etc.) draft the JSON, call APIs, and run tests on your behalf. FunnelStory supports this through the MCP server and a published flow authoring guide your assistant can read.

Human-first path

Prefer clicking the canvas? Start with Getting started—you can still paste JSON from an assistant later.

Prerequisites​

  1. Connect your assistant using MCP Getting Started and Authentication.
  2. Ensure your user can access the target workspace and connections (Slack, CRM, email).

Approach A — MCP (recommended)​

Once connected, the assistant gains tools such as:

ToolRole
query_semantic_dbExplore live schema + sample rows
get_flowsList agents or fetch a full configuration
configure_flowCreate or update an agent graph
run_flowExecute by id with optional inputs / simulated trigger payload

FunnelStory also exposes the flow authoring guide as a resource (file://flow/guide.md over MCP). In practice you can say:

“Read the flow guide, then create a draft agent that runs daily, selects accounts with health_score < 35, and posts details to Slack channel C0123 using connection <id>.”

The assistant should:

  1. Pull the guide resource.
  2. Draft trigger_config, input_schema (if needed), and config.entrypoint + steps.
  3. Call configure_flow with draft: true.
  4. Call run_flow with sample input / trigger objects to validate.
  5. Flip draft false only after you approve.

See also MCP examples.

Approach B — Skill file (offline)​

Download the same guide your workspace would expose:

Add it to your assistant’s context:

ClientSuggested location
Cursor.cursor/rules/funnelstory-agents.mdc or project rules referencing the file
ClaudeProject knowledge / uploaded doc
OtherPaste into a pinned system prompt

Keep the file version-stamped in your internal wiki if you fork it—upstream behavior evolves with the product.

Prompt snippets that work well​

  • Portfolio sweep: “List accounts where prediction = churn and subscription_remaining_days < 60. Build a query-triggered agent capped at 200 rows/day.”
  • Narrative QBR: “Given flow input account_id, pull last 90 days of meetings + tickets, then summarize risks/opps with citations in plain text.”
  • Regression test: “Fetch flow <uuid>, run it with trigger row JSON {...}, and diff the events against yesterday’s run.”

Guardrails for assistants​

  • Start draft: true until a human confirms side effects (email, CRM writes, datasets).
  • Never invent dataset names—confirm they exist via SQL.
  • Respect query-trigger daily cadence and LIMIT.

Site map for LLMs​

  • llms.txt — machine-oriented index of documentation URLs (generated at build time for the whole site).
  • llms-full.txt — single-file bundle of doc content for assistants that ingest one document.