How to use Claude to build a community eval playground
Recall built Predict largely with Claude Code subagents, then shifted from a Figma MCP design workflow to a designer prompting iterations directly into code. The author reported an end-to-end working loop with rough edges.
news.ycombinator.com
Prompt to try
Editorial adaptation, not the source author’s exact prompt. Replace bracketed inputs and review tool permissions first.
Build [community evaluation playground] with Claude Code subagents. Support proposing [skills], writing graded prompts, and forecasting [model results]. Help a designer iterate directly in code on [screens]. Separate implemented behavior from [future publication plans].
Reach people exploring AI agents
See placements →What happened
The author said Predict worked end to end, with rough edges remaining, and that direct design prompting became faster for the team.
Reported by andrewxhill on Hacker NewsHow the agent did it
- Built Predict largely with Claude Code and subagents enabled.
- Worked initially with a designer using Figma MCP.
- Shifted the designer to prompting design iterations directly into code.
Where a person stepped in
A designer and the author participated; the post does not describe code review or formal testing.
A person approved each step before the agent continued.
Andrew, a co-founder of Recall, described building Predict over several days, largely with Claude Code and its subagents enabled. Predict is a playground where people propose skills to measure in language models, write graded prompts for those skills, and forecast which models will perform best after GPT-5 is released. Examples in the post include difficult math, memory-manipulation resistance, code generation, and empathy when delivering bad news.
The implementation initially involved a designer working with Figma MCP. The team then changed its design process: the designer learned Claude Code and prompted design iterations directly into code. Andrew said this became faster for their work. The post describes the resulting loop as working end to end, while acknowledging remaining rough edges.
Predict keeps accepted evaluations private until GPT-5 ships. The stated publication plan includes prompts, responses, grades, and an on-chain checksum for provenance. These are described as product behavior and plans, rather than evidence of completed evaluation results. Who it's for: teams exploring community-authored model evaluations and direct collaboration between designers and coding agents.
Prompt
Suggested prompt (written by Agents In Use, not the author's)
Source: Hacker News · andrewxhill
Frequently asked questions
How did the designer work with Claude Code?
The team initially worked with a designer using Figma MCP. The designer then learned Claude Code and prompted design iterations directly into code, which the author said was faster.
Were the benchmark results already published?
The post says accepted evaluations remain private until GPT-5 ships, with prompts, responses, grades, and an on-chain checksum to be published afterward.
More details
About this account
Claude Code, with subagents enabled, produced most of the playground implementation while people directed the design work.
Platform: Hacker News. Attribution: https://news.ycombinator.com/item?id=44768159. Prompt: Suggested (written by Agents In Use, not the author's). Sourced by the Agents In Use pipeline.
Credit: andrewxhill on Hacker News
Source published 2 Aug 2025
Source checked 14 Oct 2026
Tools and topics
Claude Code · Figma MCP
Evaluations · Design iteration
Editorial completeness: 100/100. Measures documented fields, not agent performance or verified outcomes.