10 Parallel Claude Agents, 139 Tool Calls, Zero Code Written: AI Design Brainstorming at Scale
2 sessions. 139 tool calls. 10 design variants. And almost no time actually writing code — the majority went into figuring out what the design should even be.
TL;DR — I automated design brainstorming using Claude Code’s brainstorming skill combined with the parallel agent pattern. 10 HTML mockups were generated simultaneously, served on a local dev server, and compared in real time before committing to a direction.
Session 1: Diagnose Before You Fix
Session 1 ran 61 tool calls — mostly Read(12) + Bash(9) + Edit(16). The prompt I gave was blunt:
fix the overall design and it's broken on mobile
Instead of jumping straight into edits, Claude diagnosed first. Reading lib/content.ts and lib/types.ts revealed the actual cause of the broken archive images: the ArchiveEntry type didn’t have an image field at all, and getArchiveEntries() was silently discarding meta.image. The images weren’t “not showing” — the render path for them simply didn’t exist.
If I’d skipped this diagnostic step and jumped to “fix the design,” I would’ve restyled components without ever touching the actual bug. Reading lib/types.ts before ArchiveList.tsx made the difference.
From there it went in order: lib/types.ts → lib/content.ts → redesign ArchiveList.tsx → add warm neutral design tokens to globals.css. The agreed direction was “Editorial Tech Magazine” — warm gray palette swapped in for Tailwind’s default slate/zinc.
Session 2: The Brainstorming Skill Changes the Game
Session 2: 78 tool calls. What stands out is TaskCreate(11) + Agent(10) — that combination is where the session got interesting.
Prompt:
redesign spoonai for both mobile and web
Claude loaded the brainstorming skill and proposed Visual Companion: spin up a local server (http://localhost:54423), render mockups as HTML, and compare them directly in a browser. It started by generating 3 mood options as HTML and serving them live.
The first two rounds? “Not feeling any of these.”
all of these are bad, find a design skill
Claude loaded ui-ux-pro-max — a skill with 50+ design styles and 161 color palettes. That’s when the direction shifted. A single prompt line changed the quality of the output.
try something completely different, find 10 options, make sure images show per article
Running 10 Agents in Parallel
This is the core of the session. Claude didn’t generate mockups one by one — it dispatched 10 independent agents simultaneously.
Each agent owned a distinct design concept:
- Bento grid — soft
#f5f5f7, Apple-inspired - Masonry Pinterest — warm cream
#faf7ee+ orange accent - Neo-brutalism — hot pink
#ff5470+ electric yellow - Swiss tabular — pure white + ink
#0a0a0a - Japanese kinfolk — paper
#f7f4ee - Netflix shelf cinema — near-black
#0b0b10 - Y2K chrome retro
- Dashboard ticker — phosphor green
#22ff88
Each agent wrote its output as a standalone HTML file under .superpowers/brainstorm/. Because they ran in parallel, generating 8+ mockups took a fraction of the time a sequential run would have. TaskCreate(11) in the tool call log captures this orchestration.
The Navigation Problem (and a Quick Fix)
The parallel-generated HTML files were isolated — to view each one you had to navigate files manually.
how do i even see the different designs?
make a button at the top to cycle through them
Claude created /tmp/__nav-inject.html and injected it into the local server as a fixed top-bar navigator. One click to move between all 10 mockups in sequence. A small thing, but exactly the kind of friction that derails a comparison session if you have to solve it yourself.
Tool Call Breakdown
Out of 139 total tool calls, actual code modifications (Edit) happened 16 times — 11.5% of the session. The rest was understanding, planning, and generating mockups. TaskUpdate(27) + TaskCreate(17) = 44 calls, or 32% of the entire session, was pure agent orchestration.
| Tool | Count | Purpose |
|---|---|---|
| TaskUpdate | 27 | Progress tracking |
| Bash | 24 | Local server, git, file checks |
| Read | 24 | Code understanding, type inspection |
| TaskCreate | 17 | Parallel agent dispatch |
| Edit | 16 | Actual code changes |
| Agent | 10 | Independent design agents |
| Write | 8 | HTML mockup generation |
| ToolSearch | 5 | Skill discovery |
Three Patterns That Held Up
Diagnose before touching anything. The image bug was invisible until the code was actually read. A direct “just fix it” prompt would’ve missed ArchiveEntry entirely and landed changes in the wrong file. Reading types before components is a habit worth enforcing.
Parallel agents belong in the exploration phase. Design is exactly the kind of work where you need to evaluate multiple directions before committing. Parallel agents compress that exploration non-linearly — time saved scales with the number of variants. Once the implementation direction is locked, there’s no reason to parallelize.
Skills have to be explicitly loaded. Without ui-ux-pro-max, the output would’ve relied on Claude’s default aesthetic sense. After loading the skill, the specificity of suggestions jumped noticeably — named color values, distinct visual references, not vague adjectives. The prompt "find a design skill" was the actual leverage point for the session’s quality shift.
More projects and build logs at jidonglab.com
Comments 0