Workbench is now Playground: where should your prompts live?

According to the Claude Platform release notes:

• July 17, 2026: the legacy Workbench in the Claude Console was being sunset, with access ending on August 17, 2026. The notes say saved prompts, variables and evals are not supported in the updated Workbench, and point to an export option in the Console banner and under Organizational Settings.
• The same entry says the experimental prompt-tool endpoints (generate_prompt, improve_prompt, templatize_prompt) were to be retired with it, and return an error after removal.
• August 18, 2026: "Workbench is now playground," a tool for running Messages API requests and seeing the full SDK request and response.

According to Tech Times (August 17, 2026), legacy Workbench data could be exported as JSON until September 1, 2026, and not recovered after that. We did not find that date in Anthropic's release notes, so treat it as reported rather than confirmed. If you think you still have data there, check the Console banner, your Organizational Settings and the Claude Help Center for the current state.

AI/TLDR describes the new Playground as not storing prompts server-side. That matches the release notes' description of a request-and-response testing tool, but it is their characterisation, not Anthropic's wording.

Keeping prompt storage, version history and eval results is a different product from a model-testing sandbox. It is closer to a database with a UI. A lab narrowing its console to "try the API right now" is a reasonable decision, and most model playgrounds already work that way: good for experiments, not a system of record.

The durable lesson is independent of this one change: a provider's testing surface is not a storage system. Treating it as one makes your prompts depend on a roadmap you don't control. Take the labs' prompting advice (see what the labs agree on), but keep the prompts themselves somewhere you own.

Three options, in order of how much structure they give you:

1. Markdown files in a git repo    free, versioned, survives any vendor;
                                     no cross-surface reuse or {{variable}} filling
2. A dedicated prompt manager       versioning, search, reuse across tools;
                                     worth it past a handful of prompts
3. Your observability/eval stack    if you already run one for production LLM calls;
                                     a different job from a personal library

PromptLikeEngineer is built for the second option: prompts stay versioned in your account and come out of the CLI, the web app and MCP. Export is open on every plan (JSON, an AGENTS.md section or a Claude Skill), so you are not trading one dependency for another. That isn't a claim that we are immune to our own future product decisions; it is why the exit stays open. For the wider field, see the tools comparison. OpenAI is making a similar move with reusable prompt objects; we wrote a migration checklist.

A playground is for testing a prompt against the API right now, not for keeping it. Plan your prompt storage accordingly, whichever tool you pick.