Claude daily update: save Workbench prompts before August 17

Claude Code is still on 2.1.220. Today's practical Claude deadline sits elsewhere in Console: Workbench (legacy) closes on August 17, 2026. After that date, saved prompts, prompt versions, and evaluations will no longer be accessible there.

Source: Claude Code v2.1.220 on GitHub.

Source: Current Claude Code version in the npm registry.

Claude daily update: Workbench stops saving your work

Workbench is Anthropic's environment for trying Claude models and API requests before building them into a product. The refreshed version is stateless, which here means Anthropic does not store the prompt or conversation on its servers. The current draft stays in the browser, and you can export the tested request as code.

The legacy version could save prompts, versions, and evals. An eval is a test that measures an AI response against defined criteria. Those assets do not move into the refreshed Workbench automatically, and there is no import function.

On the same date, three experimental API endpoints for generating, improving, and templatizing prompts will stop working. Teams that call /v1/experimental/generate_prompt, /v1/experimental/improve_prompt, or /v1/experimental/templatize_prompt need to replace them before August 17.

Source: Claude Platform release notes on Workbench and the prompt tools APIs.

Human step: export before August 17

If your organization used Workbench before June 16, 2026, the legacy version may remain available until retirement.

1. Open Workbench (legacy) in Claude Console.
2. Select "Export data" in the banner.
3. Decide whether to include model completions and uploaded files. Either can make the export much larger.
4. Start the export. Anthropic packages the data as JSON and emails a download link.
5. Confirm that the file opens before treating the migration as complete.

The export can include saved prompts, prompt revisions, and completions. A Claude response appears only if it was previously added to the conversation. That detail is easy to miss.

Source: Anthropic's steps for exporting Workbench data.

Turn the export into a managed work asset

A JSON file in a downloads folder saves the content. To keep working with it, put the prompts you still use under the same version control as the workflow they direct. Connect each prompt to a purpose, an owner, example input, an accepted outcome, and a few test cases.

Anthropic's own guidance starts with specific, measurable success criteria. Test cases should match the real task and include relevant edge cases. That lets you compare a prompt or model change against something firmer than instinct.

When a prompt connects to a real system, keep its API key in an environment variable or secret manager rather than the prompt file. Use scoped permissions, an approval step for writes, and a short run log. The workflow can then do useful work while remaining reviewable.

Source: Anthropic's guide to success criteria and evaluations.

Try this prompt this week

Human step: Export the Workbench data, unpack the JSON file, and open it in a local project where Claude can read the files.

Read the Workbench export and inventory the prompts that still have a clear operational use.
Group duplicates and flag missing examples, owners, or test cases.
Propose a simple folder structure with one prompt file and test set per workflow.
Do not change anything yet. First show the three migrations that should happen first and why.
For each proposal, name dependencies, the review point, and a measurable accepted outcome.

Good output should:

• separate active prompts from historical experiments,
• show which tasks or files each prompt depends on,
• state testable criteria instead of "looks good",
• leave a reviewable next step without overwriting the export.

When several people or systems depend on the prompts, Tool Forge can turn them into a managed part of the workflow. Inventory and test them first. Integrate what passes review.