Turn a playground prototype into production code without transcription errors.
1. Finish iterating in the playground first. The playground lets you try models and API features and inspect the full response. Get the behavior right there before writing code: it is faster to iterate on parameters in the browser than in a deploy cycle.
2. Export from the code tab. The playground is built directly on the public Messages API, so use the code tab to copy the exact request as a runnable snippet in your language. Do not retype it: transcription drops parameters like temperature, top_p, or stop sequences.
3. Map every playground setting to an API field. Check the exported snippet for model, max_tokens, temperature, system instructions, messages, tools, and any beta headers, and confirm each one exists in your code path. A setting you configured by clicking but never exported is a setting production does not have.
4. Strip playground-only state. The playground does not store prompts on Anthropic's servers; your draft lives in the browser. Anything you need in production (the exact system instructions, the few-shot messages, the tool schemas) must be checked into your own repo as the version of record.
5. Re-verify in code. Run the exported request from your application and diff the response shape against what the playground showed. Differences usually mean a header or parameter did not survive the export.
6. If you used the legacy Workbench, export now. The legacy Workbench was sunset and its saved prompts, variables, and evals are not in the new playground. Export any data you want to keep from Console settings before the cutoff: after it, the data is no longer recoverable. The experimental prompt-tool APIs (generate, improve, templatize) were retired with it, so replace any automation built on them with direct Messages API calls.