Skip to main content
Job: ask for an outcome, then make Claude Code prove the play is valid before it runs.

​
Operating checklist

Before running:
  1. Confirm workspace with deepline auth status --json; for automation, run from a folder with .env.deepline.
  2. If using a prebuilt play, run deepline plays search and deepline plays describe.
  3. If authoring or editing a .play.ts, run deepline plays check <file>.
  4. Pilot with 1-3 rows or one JSON input.
  5. Inspect output and failures.
  6. Run the full job with --watch, then export the completed rows with deepline runs export <run-id> --out <output.csv>.
  7. Report input, output path, row counts, failures, and workspace.

​
Good prompt shape

Use the Deepline SDK docs.
I have leads.csv with linkedin_url and company_domain.
Write a play that finds verified work emails.
Run deepline plays check before running anything.
Pilot only the first 3 rows before the full file.
This gives Claude Code the job, the input, and the safety gate.

​
Ask for checks by name

After editing the play, run:
deepline plays check <file>
Then run:
deepline plays run <file> --csv <csv> --watch
deepline runs export <run-id> --out <output.csv>
If the play changes, check it again.

​
Keep the story visible

Ask Claude Code to summarize the result like this:
QuestionAnswer
What job did we run?Find verified work emails
What input did we use?leads.csv
How many rows succeeded?Count from output
What should a human review?Missing or uncertain rows

​
Skills are optional

Skills help agents choose better provider paths. The Deepline CLI is still the execution layer. 
/deepline-gtm Build a workflow for this CSV, then check and pilot it before the full run.
Use the skill for judgment. Use plays check for proof.