Anthropic's Latest Winner - Workbench

Anthropic has overhauled its developer “Workbench” inside the Anthropic console, turning prompt building into a full testing and benchmarking workflow. Instead of treating prompts as static text, developers can now generate prompts from scratch, run them against targeted test cases, and assemble a versioned evaluation suite to measure how well different prompt designs handle context and produce reliable outputs—then export code for use in production.

The core shift is the addition of prompt generation plus prompt testing in one place. A developer can describe a task in plain language—such as drafting a response to a YouTube comment that classifies whether the comment is toxic and whether it deserves a reply—and then click “generate prompt.” Claude produces a detailed, ready-to-run prompt that includes a system-style instruction block, the input comment, and step-by-step decision criteria. In this example, the generated instructions explicitly guide toxicity detection (offensive language, personal attacks, hate speech, threats, harassment, excessive negativity, trolling) and then determine whether a reply is warranted.

From there, Workbench moves into evaluation mode. The console prompts the user to supply the specific YouTube comment, runs the prompt, and returns a classification result along with reasoning and a final label (e.g., not toxic, moderate toxicity). Developers can generate multiple test cases automatically, run them, and then refine the prompt based on observed weaknesses. The workflow supports editing the prompt directly and re-running evaluations to see how changes affect outcomes.

A key feature is versioning. Workbench tracks prompt versions and response versions, enabling side-by-side comparisons—such as a second attempt where the developer slightly rearranges instructions (for example, moving content between user and system sections). The console can then score results across versions, letting users label performance (e.g., “excellent” or “fair”) and quickly identify which prompt variant performs better.

Workbench also streamlines what used to be a spreadsheet-and-manual-process problem. Previously, teams often exported test cases and relied on human scoring methods like MOS (mean opinion scoring) to compare prompt quality. Now, the evaluation and scoring live in the same dashboard, with the ability to import test cases from CSV and to generate additional cases on demand.

Once the evaluation suite is complete, developers can click “code” to export runnable code that reproduces the prompt set and evaluation configuration. They can also adjust runtime parameters such as maximum tokens and rerun the suite against different Anthropic models (the transcript mentions comparing Haiku and Sonnet). The result is a repeatable loop: generate prompts, test and edit them, benchmark versions, and export the workflow for trial or production deployments—directly inside Anthropic’s console.