How To Design In The Agent Era
B · Useful, limited Y Combinator Design Review episode with the founder of Paper, an HTML-based design tool; a product demo, but it gives concrete, checkable rules for removing the typical look of model-generated interfaces.
Summary based on the full transcript·Original: about 56 minutes to watch
In short
A 56-minute Y Combinator episode (uploaded 7 August 2026) with the founder of Paper, a design tool that uses HTML and CSS as its rendering engine. Because agents already read and write HTML well, a coding agent can edit a Paper canvas directly through an MCP server, and designs can be pulled back into the code base, which becomes the single source of truth. The demo reviews three founder websites and shows how to strip the "vibe-coded" look. The speaker's position: agents speed up variation, resizing and translation, but design inside an organisation is about decisions and stakeholders, and he does not expect models to take that over.
Key numbers
- threemaximum number of font sizes the speaker recommends to make generated layouts look deliberate
- about 12people in the Paper team, all designers and engineers
- about a weektime a brand designer at Paper took to design, build and ship the company website with agents
- about 25,000Twitter followers Paper gained before it had a product, by talking about its design values
As reported by the source.
English·Video·opens youtube.com
What it reports
- Rules against generated-design tells: pull font weights back from bold, use at most three font sizes, check contrast, delete decorative numbers, icons and badges, avoid all-caps kickers with wide letter spacing, purple gradients, coloured card edges, and cards everywhere.
- Paper puts these typography and contrast rules into its own instructions to models; the speaker says that is most of its "secret sauce".
- Workflow: grab a live site section as editable layers, ask an agent for several variations, leave canvas comments for the agent to address, curate, then have the agent build the chosen design into the code base. Some users leave agents generating hundreds of variations overnight.
- Model choice: Fable for visual taste but slow; Cursor's Composer 2.5 for precise bulk tasks.
- Inside Paper, people read every line of agent-written code for the core product, for performance and precision; agents are used freely for the website, videos and marketing.
- Growth came from public talk about design values and open-source shaders and a font, treated as the marketing budget.
How they got it
Product demo and founder experience. Website critiques are the speaker's judgement; no user metrics are given beyond follower counts and a reference to Ramp spending data.
Limits
- Vendor founder promoting his product on a partner channel.
- Design advice is opinion, though specific and easy to test.
- Live demos used fast models that made visible errors.
Why it matters for companies
As internal teams generate more interfaces, documents and slides with models, the generated look becomes a credibility problem: people read it as low care. The rules here are a cheap checklist for any team publishing AI-made material, and they can be written into prompts and templates. The practice of reviewing every line of agent code for the core product while letting agents run on peripheral work is a reasonable split for any firm deciding where full autonomy is acceptable.
Summary based on the full transcript. Read the original for the full argument and data.Watch the original on youtube.com
- Tags
- AI enablement, Lecture