Claude Prompt Improver: Retired, and What to Use Now

The Claude prompt improver is gone. It lived in the legacy Workbench of the Claude Console, and Anthropic retired both on August 17, 2026. Its API counterpart, /v1/experimental/improve_prompt, went too, and the prompt improver's docs page now redirects to Anthropic's general prompting guide.

You can still reproduce most of it by hand. The short version:

• What it did: rewrote an existing prompt with XML structure, step-by-step reasoning instructions, standardized examples, and a prefilled response
• What replaced the Workbench: the playground, which runs requests but doesn't save prompts, run evals, or list a prompt improver
• What still works: asking Claude to improve your prompt yourself, with the same checklist, minus the parts current models no longer need

What Anthropic's Claude prompt improver did

Anthropic announced the prompt improver on October 14, 2024, in what was then called the Anthropic Console. You pasted in a prompt you already had, and Claude rewrote it using prompt engineering techniques.

It listed five changes the improver made:

Per the same post, Anthropic's testing showed a 30% accuracy increase on a multilabel classification test and 100% word count adherence on a summarization task. Those are Anthropic's own tests, not a promise for your prompt.

How you used the prompt improver, step by step

The archived docs page laid out the flow. You needed a prompt template, plus two optional inputs: feedback on what was wrong with Claude's current outputs, and example inputs with ideal outputs.

1. Submit the prompt template.
2. Add feedback, such as "summaries are too basic for expert audiences" (the docs' own example).
3. Include example inputs and ideal outputs.
4. Review the improved prompt.

Behind that, it ran four steps you could watch live: example identification, an initial draft with XML sections, chain-of-thought refinement, and example enhancement. No examples yet? A Test Case Generator produced sample inputs and Claude's responses, which you edited into ideal outputs. You could then send more feedback and run it again.

It sat next to two related tools: structured example management in the Workbench, and an Evaluations tab with an optional "ideal output" column for grading responses on a 5-point scale. A separate prompt generator, launched in May 2024, wrote a prompt from a task description instead of improving one you had.

When it helped, and when it made prompts longer

The docs were upfront about the trade-off. The improver worked best for complex tasks that need detailed reasoning, where accuracy matters more than speed. For latency- or cost-sensitive work, Anthropic suggested simpler prompts, because improved templates produce longer, more thorough, but slower responses.

The docs' own example shows why. A short classification prompt, asking which Wikipedia article a sentence came from and to respond with just the title, came back as a template with a role, XML-wrapped inputs, a four-step ranking procedure, and an instruction to put the analysis in <analysis> tags. It then closed by telling Claude to output only the chosen title. That's more structure, more output, and two instructions pulling in different directions.

That pattern is the main risk with any automatic rewrite:

• Length becomes cost. Every reasoning section Claude writes out is output tokens, on every call.
• Reasoning gets verbose. "Chain of thought too verbose" was on the docs' troubleshooting list. The fix was adding length instructions back in.
• Steps are generic. Another listed issue: reasoning steps that don't match your use case. The improver guessed at a procedure, and you corrected it.
• Missing information stays missing. If your prompt lacked source material or a clear goal, the rewrite had the same gap. Our look at AI prompt enhancers covers that failure mode across tools.

Why old improved prompts don't fit current Claude models

If you saved improved prompts, some of what the improver added now works against you.

Prefills break. Anthropic's prompting best practices say prefilled responses on the last assistant turn aren't supported starting with Claude 4.6 models. Those requests return a 400 error. The improver's "prefill addition" step is exactly that. Move the prefill's job into the instructions, for example "respond with only the JSON object."

Reasoning is built in. Claude 4.6 and later models use adaptive thinking, where the model decides when and how much to think. On Fable 5 and Fable 5.1, thinking is always on. A written reasoning section asks for visible reasoning on top of that, which is often just extra output.

Aggressive wording overshoots. The same guide says Claude Opus 4.5 and Opus 4.6 are more responsive to the system prompt than earlier models, and suggests dialing back language like "CRITICAL: You MUST." Heavy templates tuned for older models can now overtrigger.

Response length varies by model. Anthropic notes Claude Opus 5's default responses run longer than earlier models', and recommends asking for conciseness explicitly. For the speed and cost side of model choice, see Claude Haiku vs Sonnet.

What happened to the Workbench and its prompt tools

Anthropic's release notes cover it in two entries. On July 17, 2026, they announced the legacy Workbench would be sunset, with access ending August 17, 2026. The experimental APIs for generating, improving, and templatizing prompts (/v1/experimental/generate_prompt, /v1/experimental/improve_prompt, and /v1/experimental/templatize_prompt) were retired with it, and return an error after removal. On August 18, 2026, the notes say Workbench is now playground.

Anthropic's playground help article lists what the playground does: run requests on the Messages API, switch models and settings, test tools and structured outputs, view the raw request and response, and export the request as code. It doesn't save prompt history or evaluate prompts. Your draft stays in your browser. A prompt improver isn't among the listed features.

If you had saved Workbench prompts, the help article set an export deadline of September 1, 2026. After that date, the data isn't recoverable.

How to improve a Claude prompt yourself now

You can get most of the improver's value by asking Claude directly, in the playground or in a normal chat. The difference: you decide which changes to keep.

1. Start with the failure, not the prompt. Run your current prompt on 3 to 5 real inputs and note what's wrong. That was the old "feedback" field, and it's the input that matters most.
2. Write ideal outputs. For two or three of those inputs, write the answer you actually wanted. Anthropic's guide recommends 3 to 5 examples, wrapped in <example> tags.
3. Ask Claude for a constrained rewrite. Something like this:

Improve the prompt below. Problems with current outputs: [paste your notes]

Rules for the rewrite:
- Separate instructions, context, and inputs with XML tags
- Keep my examples, formatted consistently inside example tags
- State the output format explicitly
- Do not add a prefilled response
- Only add a step-by-step reasoning section if the task needs multistep reasoning
- Do not invent requirements I did not state; list any assumptions separately
- Keep it as short as possible while fixing the problems

Prompt:
[paste your prompt]

4. Test old and new side by side on the same inputs. Without an Evaluations tab, a spreadsheet works: input, old output, new output, which is better. If the rewrite doesn't win, keep the original.
5. Save the winner somewhere you control. The playground stores nothing, so versioning is on you.

Starting from a blank page instead? Anthropic's metaprompt notebook generates a prompt template from a task description. The notebook itself says it's built for single-turn prompts and the result isn't guaranteed to be optimal. For prompts your whole team keeps refining, see how to build a prompt library people actually use. For techniques past structure and examples, try advanced prompt engineering techniques.

A prompt that finally works is easy to lose, especially now that the playground keeps nothing. Taku is an AI-native desktop workspace where you can mirror an AI setup someone already got working, run it on your own tasks, remix it, and keep it instead of rebuilding it from an old chat. Taku is in Beta, and the Mac app is available now.

Key Points

• The Claude prompt improver was part of the legacy Workbench, which Anthropic retired on August 17, 2026, along with the improve_prompt API.
• Anthropic's docs said improved prompts produce longer, slower responses, so it suited accuracy-first tasks.
• Prefills no longer work on Claude 4.6 and later models, so old improved prompts may need editing.
• The playground replaced the Workbench, without saved prompts, evals, or a listed prompt improver.
• You can reproduce the improver by asking Claude for a constrained rewrite and testing it against the original.

FAQ

Does Claude still have a prompt improver?

Not in the Claude Console. The prompt improver lived in the legacy Workbench, which Anthropic retired on August 17, 2026. Anthropic's help article for the playground, its replacement, doesn't list a prompt improver among its features. You can still ask Claude in any chat to improve a prompt.

Can I still call the improve_prompt API?

No. Anthropic's release notes say the experimental generate_prompt, improve_prompt, and templatize_prompt endpoints were retired with the Workbench on August 17, 2026, and requests to them return an error.

Did the Claude prompt improver make prompts longer?

Yes, by design. Anthropic's docs said improved templates produce longer, more thorough, but slower responses, and suggested simpler prompts for latency- or cost-sensitive applications.

How do I improve a prompt for Claude without the Console tool?

Collect real outputs and what's wrong with them, write a few ideal answers, then ask Claude to rewrite the prompt under explicit rules: XML structure, a stated output format, no prefill, and no invented requirements. Test both versions on the same inputs.

What happened to my saved Workbench prompts?

Anthropic offered a JSON export from Console settings until September 1, 2026. Its help article says the data isn't recoverable after that date, and the playground doesn't store prompts on Anthropic's servers.