Automate the workflow nobody wants to do

Start with the task someone repeats weekly, that follows rules, and that fails quietly when skipped.

Pick the first one

• It has a clear trigger and a clear finished state.
• The inputs are available digitally.
• A mistake is cheap to catch and fix.

Build that one end to end before adding a second. A single workflow that runs unattended teaches more than five half-built ones.

Use durable execution for anything that waits

Real workflows wait: on a provider to finish processing, on a rate limit, on a human. Running them as durable workflows means a restart doesn’t lose the run, retries happen without duplicating work, and a long pause is just a state.

Keep the data where it belongs

• Workflow state, provider IDs and approvals in the operational database.
• Personal contact data in that database only, never in analytics or logs.
• Reporting data in object storage, queried through views.

Keeping these apart makes it possible to answer “where is this stored” for every field.

Test the path that matters

Offline tests cover logic, validation and HTTP behavior with fakes. One live test runs the real delivery flow end to end and reads back what the providers hold. Don’t cover the same behavior twice.