Choose what you want to make. Each guide gives you a useful starting prompt and the evidence to ask for before the work is complete.
◎
7 minute guide
How to build a headless agent
Create an agent that runs from an API call, webhook, schedule, inbox event, or chat without keeping a browser open.
Start with this prompt
Build a headless agent that qualifies inbound leads.
It should read the lead and our qualification rules, enrich the company, score the opportunity, update the CRM, and draft a reply. Require a human approval before sending. Expose a callable endpoint and keep a visible run history.
Name the trigger and outcome.Say what starts the work and what a successful result looks like.
Give the agent tools and knowledge.Connect only the records, APIs, and guidance it needs.
Mark the human boundary.State which actions need approval and which can run unattended.
Verify the endpoint and failure path.Run a safe example, a rejected request, and an approval handoff on Preview.
Build the interface and the platform behind it in the same working Preview.
Start with this prompt
Build a customer portal for our service business.
Customers can sign in, upload documents, see invoices, and ask a question. Owners can review every account and receive a weekly summary. Make it work on mobile and use our existing brand.
Describe the user journey.Name who uses it, what they do, and what they should see next.
State the private data rules.Say which records belong to each customer and what owners can access.
Build on Preview.Keep Published untouched while the interface and Wok resources change together.
Test the real journey.Use the live Preview URL for sign-in, upload, data, email, and mobile checks.
Run recurring work with retries, useful alerts, and evidence that it finished.
Start with this prompt
Every weekday at 8:00, review new applications, flag missing documents, and email the owner a concise summary.
Do not send the same summary twice. Record each run and alert the owner only when the job fails after retrying.
Set the schedule and timezone.Use an exact cadence and say when late work should catch up.
Make the work idempotent.Give repeated runs a safe way to recognize completed work.
Define useful failure behavior.Set retries, the final alert, and the evidence that helps diagnose it.
Run a safe test.Trigger it on Preview and verify the record, output, and next scheduled run.
Expose focused tools to coding agents and other MCP clients through a production endpoint.
Start with this prompt
Build a Streamable HTTP MCP server for our account data.
Add read-only tools for account lookup, invoice status, and recent activity. Authenticate every caller, enforce account access inside each tool, validate inputs, and keep an audit record. Return tested client configuration.
Design narrow tools.Give each tool one clear job and a bounded response.
Enforce access at the resource.Never trust scope or identifiers supplied by the calling agent.
Return stable failures.Validate inputs and give callers actionable error codes without private detail.
Test as an MCP client.Verify discovery, an allowed call, a denied call, and the final Streamable HTTP address.
