Miju Labs

← All buyers · Anthropic

P

Prithvi Rajasekaran

Anthropic Labs — engineer/researcher (author, harness design for long-running apps) · Anthropic

medium confidenceTier A · Talk to firstFrontier labsUI

The Anthropic Labs engineer who wrote the company's public design grader — a four-dimension rubric (design quality, originality, craft, functionality) used to score the output of long-running app-building agents. He also co-wrote the November 2025 piece diagnosing why model-generated UI looks generic. He is the closest thing in the industry to a named owner of 'how we score whether a model has taste', and his own write-up concedes the grader is calibrated on few-shot examples rather than on collected human preference data.

Background

He is a member of the Labs team and, earlier, of Anthropic's Applied AI team, where he co-authored 'Improving frontend design through Skills' (12 Nov 2025) with Justin Wei and Alexander Bricken. That piece named the problem precisely: LLMs converge on the safe centre of their training distribution, producing Inter fonts, purple gradients on white, and no motion — output users instantly read as machine-made. The remedy they shipped was a Skill that forces explicit choices across typography, colour and theme, motion, and backgrounds.

The March 2026 harness-design post is the sequel and the more important document. Building an agent that can run for hours on a full-stack app, he needed a generator-evaluator loop, and therefore needed something to evaluate design. He wrote a grader with four weighted dimensions — design quality (does it feel like a coherent whole rather than a collection of parts), originality (custom decisions versus template defaults and AI patterns), craft (typography, spacing, colour harmony, contrast) and functionality (usability independent of aesthetics) — weighting the first two more heavily. Calibration was by few-shot examples with detailed score breakdowns, aligned to designer preference by hand. There is no mention in the post of human raters, a preference dataset, or a held-out eval set.

What they run now

Agent harness design for long-running autonomous app building: a three-agent generator/evaluator structure where a design grader and a functionality grader score the artefact and drive iteration. His practical problem is grader reliability — an LLM judge scoring visual craft with no ground truth beneath it.

What they buy

This is the most concrete buying signal in the batch. He needs, and does not have: (1) a human-graded calibration set for his four dimensions — professional designers scoring generated UIs with written rationale, so the LLM grader can be validated rather than asserted; (2) disagreement data showing where the LLM judge and designers diverge, which is what turns a rubric into a reward model; (3) originality labels — 'is this a template default or a real decision' is exactly the judgement a working designer can make and a model cannot. Inference from the post, but the gap is stated in the post itself.

Career

  1. 2025Applied AI team, Anthropic — Frontend-design Skills work
  2. presentLabs, Anthropic — Harness design for long-running apps

On the record

Design grading rubric · 2026-03

A good design should feel like a coherent whole rather than a collection of parts; originality means custom decisions rather than template defaults. source

Rubric weighting · 2026-03

Design quality and originality are weighted more heavily than craft and functionality. source

Generic AI UI · 2025-11

Models converge on safe universal choices because those dominate the training data — Inter fonts, purple gradients, minimal animation. source

Interviews & talks

Why it matters here

Very high — probably the highest technical-fit target in the batch. He has published the rubric, so Julian does not have to guess the spec; he has publicly conceded the calibration weakness; and the artefact he owns (an agent that ships apps) has an obvious business reason to be better. A vendor that arrives with the four dimensions already operationalised and a pilot dataset scored against them is speaking his language exactly.

How to reach

Direct and technical. Reproduce his grader, run it against a panel of professional designers on a few hundred generated interfaces, and publish the agreement statistics per dimension. Lead with the number — 'your grader and our designers agree on craft but diverge on originality' — and offer the calibration set. Do not lead with capacity or pricing. Speculation: as a named blog author he is reachable via the Anthropic engineering blog's channels and via the claude-code plugin repo where the frontend-design Skill lives.

What we could not establish
  • No X handle, LinkedIn or personal site found.
  • Exact job title (engineer vs research engineer vs member of technical staff) not evidenced.
  • Whether the design grader was ever validated against human raters after the post.
  • Whether he or a central Human Data team would commission such a dataset.

Why they are on the map

Wrote Anthropic's frontend-design grading rubric (design quality, originality, craft, functionality) for long-running app agents; co-authored the frontend-design Skill post (Nov 2025). source — evidence dated 2026-03

Needs the data for a model; writes the spec.

How Anthropic buys data

Open roles (updated Aug 2026) for 'Data Operations Manager, Human Data' (owns data strategy and 'strategic vendor partnerships', $270k-365k) and 'Product Management, Human Data Platform' (labeling interfaces, vendor tooling, crowdworker sessions, data-quality KPIs). The Information (via TechCrunch, Sep 2025) reported Anthropic leaders discussed spending over $1B on RL environments, working with Mercor and reportedly Mechanize. Surge AI publicly marketed that its human data powers Claude 3 (2024). Anthropic Labs' own design grader (Mar 2026) scores design quality, originality, craft and functionality, which is the closest public sign of design-taste evaluation.

Human Data sits alongside research: Data Operations Managers own data strategy and vendors for research teams (RLHF, safety, tool use, agentic), and a Human Data Platform team builds labeling interfaces, environment simulation and vendor tooling. RL environments have their own infra team ('Environments Infrastructure'). Claude Design is built inside Anthropic Labs (Krieger/Mann), and Jenny Wen leads design for Claude.

Known vendors: surge-ai, mercor, mechanize

Sources

  1. Harness design for long-running applications
  2. Improving frontend design through Skills
  3. Anthropic Designs Three-Agent Harness
  4. frontend-design plugin