Same agent. Same prompt. The only difference: rules.

method

Two arms, everything else held constant

Round 1: build the dashboard (metric cards, sortable table, settings form). Round 2: wire it to a real Express backend with loading and error states — the round that exercises async rules.

results — what the counters say

After round 2 (API-wired build)

Read the last row carefully — it's the honest part. Cursor already avoids any and empty catches by default; rules matching strong model defaults added nothing. The rules that changed behavior were the ones encoding opinions the model doesn't have: your palette, your timeout policy, your taste.

results — what the screenshots say

Arm A: the slop is the point

arm a — no rules

arm b — with rules

Arm A is what you get when an agent averages the internet: indigo-to-violet gradients, four identical metric cards, a money-bag emoji, "Welcome back, Alex. Here's what's happening with your workspace today," and a generic brand name ("Nexus"). Every dashboard from every agent looks like this, because it's the statistical average of every dashboard ever published.

Arm B is the same model, same prompt — with 30 lines of rules in context. Charcoal surfaces, one emerald accent, asymmetric grid, monospace numbers, flat borders. It looks like a team made it on purpose.

mobile — 390px, where slop hurts most

arm a — no rules

arm b — with rules

Same pattern at 390px: Arm A's four-card row and wide table compress into unreadable strips. Arm B's rules included responsive behavior, so the layout stacks instead of shrinking.

in the code

The diff you'd review

// Arm A — fetch with no timeout: hangs forever on a stuck server
const res = await fetch("/api/metrics");

// Arm B — the rule, applied
const res = await fetch("/api/metrics", {
  signal: AbortSignal.timeout(10_000),
});

// Arm A — the palette
background: "linear-gradient(135deg, #6366f1, #8b5cf6)"

// Arm B — the palette
background: "#17181b", border: "1px solid #26272b"

The rules file is 30 lines. That's the whole intervention. Read it in the experiment repo: github.com/Amaresh/deslop.