S01 · E04 — THE DESIGN AGENT16:9 · ANAMORPHIC · BLEACH

TheDesignAgent

The judgment layerfor agent-built UI.

JTBD-calibrated UX and visual evaluation for agent-built UI. Your agent calls /discover before the build, /ux and /visual after. Findings cite the job — not a generic rubric. 5 free credits at signup. No card.

start free →see how it works ↓

VHS · TRACKING OKTIMECODE LOCKED · DOLBY SR

◢ 04THE GAP YOU KEEP HITTING

Your agent ships the wrong UI — the same three ways, every time.

Not a benchmark — the moves your agent reaches for when it doesn't know the job. You've fixed all three this week.

◢ 01 · DEFAULT

“It built another dashboard.”

— when the user just needed to do the thing.

The job was do the thing, not watch the thing. Your agent didn't know the difference.

◢ 02 · DEFAULT

“Tabs again.”

— even though the job is sequential.

High-stakes, sequential workflows split across status tabs. The default move when the agent doesn't know the flow.

◢ 03 · DEFAULT

“The shadcn defaults are showing.”

— and the diff passed review.

The UI ignores who's using it, when, and why. It looks fine. It's wrong. The user can't get through it.

◢ 05RECEIPTS · BEFORE / AFTER

Here's the same agent, the same model — with us in the loop.

Identical prompt. Identical model. The only difference is the brief before the build and the scoring after. Drag the slider to see what the judgment layer caught.

PR Brief · #847

opened 14m ago · auto-drafted

Fix race in webhook delivery dispatch loop

branch
feat/webhook-race-fix → main
author
agent / claude-4-opus
diff
+127 / −73 · 6 files

Files · 6

by risk

src/webhooks/delivery.ts

payments

+47·−23

src/webhooks/queue.ts

+28·−14

src/utils/mutex.ts

shared

+18·−12

src/webhooks/types.ts

+6·−0

tests/webhooks/delivery.test.ts

+tests

+24·−0

/CHANGELOG.md

+4·−0

Diff shape

+127200 LOC touched−73

Risk score

6.4/ 10high

▼

lowmedhighcrit

Replaces ad-hoc setTimeout retry with a two-lock dispatcher. Closes the duplicate-charge race seen in INC-2418. Touches payment-critical code in 1 of 6 files.

coverage
+2.1%
bench p99
−4 ms
deps
0 added
migrations
none

src/webhooks/delivery.ts · hunk 1 of 3·  the line concern #1 points at
141 async dispatch(event: WebhookEvent) {
142+  await this.queueLock.acquire();
143+  await this.replayLock.acquire();← lock-order risk
144   try {
145     return await this.deliver(event);
146   } finally {
147+    this.replayLock.release();
148+    this.queueLock.release();
149   }

AI-flagged concerns · 3

self-reported by agent

01

Lock order can deadlock against the retry worker

New dispatch acquires queueLock → replayLock. The retry worker at queue.ts:88 acquires them in the opposite order. Concurrent fire is reachable from the hourly cron.

delivery.ts:143–149

02

Test only covers the single-dispatcher happy path

No test simulates two workers dispatching the same event in parallel — which is the case this fix is intended to address.

delivery.test.ts:14–42

03

WebhookRetryAdapter deleted; downstream not checked

In-repo grep finds no callers. External services (billing-cron, ops-dash) were not scanned by this agent.

queue.ts:201

Estimated read

3min

vs. budget

−1m under 2m goal

Verify before ship

0/3

Read mutex.ts diff — 8 lines

Replay one Stripe + one GitHub fixture locally

Confirm no external consumers of WebhookRetryAdapter

Agent record · last 5

4 / 5 shipped clean

#841

rate-limit headers

#839

idempotency keys

#836

first pass missed RLS

#832

flake in prod

#828

retry backoff

Before

Slot · after

PR review · with TheDesignAgent

Paste the built component into this slot.

After

◢ 05TONIGHT'S SEGMENTS · THREE TOOLS

One MCP. Three tools. Calibrated to the job — not a generic rubric.

/discover before the build, /ux after, /visual when the screenshot matters. Same agent. Same model. New behavior.

◢ SEG · 01

/discover

Reads your system first.

— a JTBD brief, before the agent draws a pixel —

JTBD — who's using it, what's at stake.
Heuristics — picked for this job.
Patterns — surface, density, rhythm.
Project model — tokens, components. Cached.

◢ SEG · 02

/ux

Scores against the brief.

— findings cite the job + the file:line —

Heuristic check — Nielsen, JTBD-shaped.
Cognitive load — density vs. the work.
Pattern fit — graded vs. Discover.
Findings — file:line + the fix.

◢ SEG · 03

/visual

What code review can't see.

— headless screenshot, scored vs. your tokens —

Brand fit — palette, type, hierarchy.
Density — rhythm vs. JTBD profile.
Contrast — AA at ship level.
Diff — vs. last passing run.

◢ 06WHO'S TUNED IN

Built for whoever owns the UI.

Four roles. Same gap. One install closes it — without changing your agent, your stack, or your taste.

◢ ROLE · 01

Builders

Ship right the first time.

— for the builder who's redone agent UI three times this week —

Same agent, same model — new behavior, no new tool.
Brief before pixels — score after. Findings cite file:line.
One install — every repo, every editor.

◢ ROLE · 02

Forward Deployed

Drop in any repo. Any stack.

— custom UI in customer code, every week —

No infra — one pointer file. Stdio MCP.
Reads tokens — ships in the customer's voice.
file:line + fix — agent applies before opening the PR.

◢ ROLE · 03

Designers

Your taste becomes the constraint.

— stop watching your system flatten into shadcn defaults —

Tokens enforced — palette, type, hierarchy at the pixel.
Pattern fit — graded against your brief, not a generic rubric.
The agent — has to pass through you.

◢ ROLE · 04

Product Managers

Out of the QA loop for agent slop.

— every build scored before it reaches you —

JTBD-calibrated — findings cite the actual job.
file:line + the fix — agent applies before the PR.
Reviewable — every run a diff. Nothing AI-slop lands.

◢ MIDROLL · DISPATCH FROM THE EDITOR

AI-generated UI looks AI-generated.
That's a bug — not a vibe.

The Manifesto · entry 01 of 05
read the full transmission with $ cat manifesto

◢ 11FROM THE FIELD NOTES

What we're learning while building this.

◢ Concepts

The component library was the right answer to a problem agents no longer have

Component libraries exist because building UI was expensive and inconsistency was hard to prevent. AI agents changed the cost equation. The case for building unencumbered — and what you need instead.

5 min read →◢ Anti-patterns

The dashboard default: why agents build viewers when the job is deciding

Four KPI cards, a table, a date filter. The dashboard is the agent's lowest-effort answer to 'build me a tool for this data.' It's also wrong most of the time.

4 min read →

all field notes →

◢ INTERSTITIAL · 02:14

Your agent will not become a designer.It can ship like one is in the loop.

◢ 12SUBSCRIPTION PACKAGES · BILLED MONTHLY · CANCEL ANYTIME

Five credits to start. A subscription that pays for itself.

Every plan drops credits in monthly and resets at the end of the cycle. Bonus credits roll over and never expire — even after cancellation. We don't charge for errors, and you pay half on degraded calls.

5free credits

Start free. No card.

Five credits drop in at signup — enough for 1 /discover + 1 /ux + 1 /visual with one credit to spare. Use them in any repo, any agent. When they run out, pick a channel below.

no credit cardinstall in 60scredits never expire

start free →

1credit · /discoverproject model + JTBD brief, cached after first call

2credits · /ux4-stage JTBD evaluation with file:line findings

3credits · /visualheadless screenshot + token-fit scoring

CH · 01 — Starter

$27/mo

30 credits / month · +3 bonus rolls over