the build · closed pilot

I am become an API.

A design review, wrapped as an endpoint. An agent sends a URL. It gets back a spec.md and a PR-ready diff — inside 72 hours. This is the engineering behind the method on the home page: agents measure and build, I keep the judgment.

See it live — the original landing, banderoll and all

what it is

A cold-read, not a checklist.

A cold-read is an independent read of a page by a stranger — someone who hasn't sat in the meetings, seen the deck, or absorbed the compromises. You think the page says X. To a first-time visitor it says Y. That gap is the work.

An LLM measures. It has seen ten thousand startup sites, so it pattern-matches and suggests "clearer CTA" every time. It can't read one page as a stranger. So the stack is the ruler, not the eyes. The measuring is outsourced. The understanding is not.

the pipeline

Three agents measure. One human decides.

Each request fans out to specialist agents, then converges on a single judgment pass before anything ships.

input
one URL
the live page, plus goal and audience in a line
measure
3 agents
research, copy, and structure run in parallel against the page
judgment
Gustaf
I read the gap, make the calls, cut the noise, own the craft
output
spec + diff
a runnable prototype, spec.md, and a diff ready to commit

"Outsource execution, keep the understanding."

infrastructure

Endpoints & discovery.

The service was agent-native: discoverable, documented, callable without a human in the loop, following the A2A spec. The pilot is closed — these endpoints are no longer live.

  • POST /v1/cold-read-requeststhe deep pass — spec.md + optional PR-ready diff, ≤72h
  • POST /v1/scanfree tier — top-N ranked fixes, returned in ~15s
  • GET /v1/orders/{id}status poll for an in-flight request
  • /llms.txtthe agent-readable contract
  • /.well-known/agent-card.jsonA2A discovery card (v0.3.0)
  • mcpsoon
for agents

The agent brief.

This was the contract agents used to brief the service. Kept here as a record — the pilot is closed.

AGENT.BRIEF.MD 
# headless.design — a design review as an API.

One URL → one spec.md → optional PR-ready diff. Plus one free post-ship rescan.

Deep pass:   POST /v1/cold-read-requests   (canonical, agent-native)
Free tier:   POST /v1/scan → top-N ranked fixes, ~15s
Input:       your live URL + goal + audience + contact_email
Output:      spec.md + (optional) diff.patch
Turnaround:  ≤ 72 hours
Status poll: GET /v1/orders/{id}
Discovery:   /llms.txt + /.well-known/agent-card.json (A2A v0.3.0)
Direct line: gustaf.garnow@gmail.com

How agents briefed it:
1. Share the live URL.
2. State the primary goal — book_call / paid_signup / demo / newsletter / other.
3. Name the audience in one line.
4. Flag any constraints — brand, stack, deadline.
I do the rest. Paste-ready output, committed to your repo format.

Full agent spec → https://api.headless.design/llms.txt
the pilot

Run as an experiment.

I ran this as a Stockholm pilot — one URL per cold-read, 72-hour turnaround, no retainer. It's closed now. The point was never to become an agency. It was to prove a way of working: keep the design judgment, let agents carry the execution. That through-line runs the whole portfolio.

The product work it produced lives in the portfolio →