Prompt-craft & iterating with Cascade
Goal. Drive Cascade conversationally — refine your script by describing precise, located changes instead of restarting.
Add a command-line argument for the input filename so the script isn't hard-coded. Then filter out any hit with an e-value above 1e-5, and print how many queries were kept versus dropped.
Send this as a follow-up in the same Cascade conversation — it keeps iterating on the code it already wrote instead of starting over.
- 1Keep each request specific and located: name what to change and where ("add a CLI argument for the input filename", "filter hits above 1e-5") rather than "make it better".
- 2Cascade stays aware of your project and the terminal, so it can pick up mid-task — group a few related tweaks into one message and let it apply them together.
- 3Watch the proposed edits, accept them, and re-run. Iterate in small, focused batches; if a request was too vague, restate it more precisely rather than piling on.
You'll see. Your existing script updated in place with the requested changes, the earlier work preserved.
Cost. Tab completions and inline edits are unlimited and free on every plan; each Cascade task draws on your agent quota, so batch related changes to make them count.
Takeaway. You steer Cascade like a teammate: describe outcomes precisely, iterate in small steps, and let it keep building on what it already wrote.