СШАAnthropic
Product Designer, Evals & Prompts
1 мес. назад
305k–385k USD / годСШАSeniorHybrid
pythonevaluation pipelinesgradersrubricsa/b testingfront-endsandboxing
Anthropic is an AI safety and research company building reliable, interpretable AI systems like large language models. The role focuses on designing and evaluating prompts and evaluation systems to ensure alignment of AI products with user expectations and safety requirements.
About the company
- Anthropic’s mission is to create reliable, interpretable, and steerable AI systems.
- We want AI to be safe and beneficial for our users and for society as a whole.
- Our team is a quickly growing group of committed researchers, engineers, policy experts, and business leaders working together to build beneficial AI systems.
About the product
- Prompts specify what Claude does.
- Evals measure whether it did.
- The Product Prompt and Eval Design team works on both prompt creation and evaluation for product design including system prompts for new users, tool descriptions, instructions, and evaluation tests.
- The goal is to keep the model and product aligned with user expectations, product strategy, and safety on every surface and model launch.
- This role builds evals that check prompts, the harness that runs them, and tools that allow designers to work autonomously.
Responsibilities
- Write and revise the prompts behind Claude's tools, features, and behaviors on a product surface; test the surface, turn findings into prompt fixes, ship them, and confirm the prompt users get is the intended one.
- Build the graders that prove a prompt fix, rerun on the next model; turn designers' hand-run rubrics into automated evals, then read transcripts for what the eval missed.
- Build visual, low-code eval tools designers can use without an engineer: assemble comparison sets from real transcripts, turn plain-English rubrics into graders, compare prompt variants across models side by side, and read results in the tool vs a notebook.
- Observe designers using these tools and simplify them.
- Support model releases by testing each surface against new models, writing prompt fixes, migrations, and prompts for new features with data backing.
- Stand up and scale the eval harness, building a test environment for 50 to 100 tools, keeping evals consistent across models, and diagnosing regression sources.
- Package unfixable prompting issues for training with attached evals: graders for crisp behaviors, human-feedback questions, and good/bad pairs for fuzzy behaviors.
Requirements
- Production-quality Python.
- Experience building and maintaining evaluation pipelines for LLM products: graders, rubrics, comparison sets, regression suites, and plumbing for cross-model execution.
- Experience building internal tools with real interfaces for non-coders.
- Experience setting up test harnesses, sandboxing tool calls, and pinning settings for comparability.
- Experience shipping prompts or working closely with prompt engineers, understanding prompt failures across models.
- Ability to read transcripts, not just scores.
- Minimum education: Bachelor’s degree or equivalent.
- Required field of study relevant to the role.
- Visa sponsorship support available but varies by role and candidate.
Nice to have
- Experience within a model-launch cycle.
- A/B testing experience linking offline evals to online outcomes.
- Front-end or notebook-to-app experience and opinions on eval result legibility.
- Experience converting product rubrics into training signals like graders, feedback questions, or preference pairs.
- Care about Claude's behavior for users, beyond metric movements.
Conditions
- Hybrid work model with offices in San Francisco or New York.
- Expected to be onsite at least 25% of the time.
- Visa sponsorship is possible with support from immigration lawyers.