AgentAgents I use on this site

An AI web design agent, and the separate reviewer agent that checks it

Two agents split the work. The design agent starts with the visitor, writes a short concept, tries three directions, builds one, and checks it in a browser at several screen sizes in light and dark. A second agent, which can read but never edit, is set up to critique the result, because a reviewer has no reason to defend the design. On this site that reviewer could not be launched, so the design agent checked its own work.

Go to the agent brief

Hard limits

Written into the agent's own brief.

  • Never commit or push.
  • Build on the site's design tokens only. Do not invent facts, numbers or logos.
  • Read-only. It cannot edit files.
1

Copy the agent brief

Nothing to fill in. Copy it as it is.

What the agents are told to do11 lines
Design agent- Think first: who lands here, what do they want in five seconds, what makes them doubt, what is the one next step.- Write a one-sentence concept, sketch three different directions, pick one and say why.- Build on the site's design tokens only. Do not invent facts, numbers or logos.- Check in a browser at 375, 768, 1280, 1440 and 1920 pixels, in light and dark.- Never commit or push. Design reviewer- Read-only. It cannot edit files.- Opens the page in a real browser and scores it against the visitor's goal, an anti-generic checklist, accessibility and truth rules.- Returns a ranked list of fixes.

How to use it

  1. 1Write down the visitor and the one next step before you ask for any design.
  2. 2Ask the design agent for a concept and three directions first, then approve one.
  3. 3Let it build, then run the reviewer on the result. If the reviewer cannot run, as happened here, treat the design agent’s own check as the weaker one and look at the page yourself.
  4. 4Fix the reviewer’s top items, and check the page yourself on a real phone.

A good result looks like

A page that starts from a clear idea about the visitor, works at every screen size, and comes with a critique that names its weak points.

Stay safe

  • Agents can miss things a person sees. Look at the result on real devices.
  • Keep the reviewer read-only so it cannot change what it is judging.

If it goes wrong

4 known problems, each with a fix

  1. 01

    Problem: The design looks like every other AI-made page.

    Fix: The concept step and a list of banned patterns are there to prevent it. Reject the concept if it could fit any site.

  2. 02

    Problem: The agent changes copy that should stay.

    Fix: Tell it the copy is final and the job is visual only.

  3. 03

    Problem: Browser checks are skipped.

    Fix: Ask for a list of exactly which sizes and themes were opened.

  4. 04

    Problem: The page looks right but a fact on it is wrong.

    Fix: The agent’s rules say not to invent numbers or logos, but still check every claim against your own source.

Where I have used it

  • Used on this siteClaude Code subagent · 22 Nov 2026

    Used on this site to redesign the AI Lab pages. The reviewer agent could not be launched in that session, so the review step was run by the design agent itself and is noted as a limit.

Questions

Why two agents instead of one?

In my view a builder tends to favor its own work. A separate reviewer with no power to edit should give a more honest read and make the weak points visible.

Did an agent build the AI Lab pages?

Yes. On 2026-11-22 an agent redesigned the AI Lab pages. The agent’s own report lists what it checked in a browser and what it did not, and the review was done by the design agent itself, not by the separate reviewer.

What is an AI design agent?

It is an AI assistant given one job, a fixed method and hard limits, instead of a blank chat. Mine works inside Claude Code and follows this site’s design rules, so it does not invent a new style each time.

Can you build something like this for my product?

I build AI integrations, bots and automations. If you want an agent that does one defined job in your own workflow, with limits and a person approving the risky steps, describe it on the estimate page.

Build this for you

Want this built into your own workflow?

Tell me what you do by hand today. I will say what is safe to automate, what needs a person, and roughly what it would take. I reply within 24 hours.

Wondering what your project will cost?

Get a free price range and timeline in 24 to 48 hours. No commitment.

Get a Free Estimate