Pimp My IDE / Garage dispatch
Back to garage
September 28, 2026 | prototypes / critique / taste

The prompt is not the product. The next visible draft is.

Parallel agents can make more options than one person can inspect. The advantage appears only when each option becomes a named preview, a focused critique, and a tighter next brief.

The take. Do not ask for one perfect design. Ask for distinct candidates. Put them on the same inspection surface. Keep one. Name why. Send the next lap toward that reason.
Run a critique lapRead the field report

Volume is not the method.

Felix Rieseberg describes a homepage redesign that used about 60 threads. Ten threads explored early directions. The shipped version took about 60 iterations. Those numbers are striking, but the repeatable part is smaller. Each thread published a private preview. Rieseberg inspected options, chose a direction, and sent back critique.[1]

The loop converted an opinion into another artifact. "More creative" was not the finish line. A visible candidate gave the next request something concrete to preserve or reject.

Parallel work buys search width. Critique decides whether any of it climbs.

Name the candidate before judging it.

A pile of tabs becomes hard to discuss. Give every direction a short name and one sentence that states its governing idea. Then compare the same properties across all candidates: silhouette, reading order, narrow-screen behavior, and cost.

This does not require a numerical design score. It requires stable language. "Night desk keeps the single-room frame" is useful. "Make it better" is not.

Generic warnings produce generic swaps.

Anthropic's Opus 5.5 prompting guide says a broad instruction such as "avoid a generic AI look" tends to replace one default style with another. The guide recommends naming specific patterns to avoid, checking the first result, and extending the instruction from what appeared.[2]

That is a critique loop, not a magic phrase. Record the unwanted pattern you saw. Record the property worth keeping. Make both part of the next brief.

Keep sessions separate and previews cheap.

VS Code 1.139 stores lightweight session and chat metadata in a central catalog while keeping full conversation content in each session database. It also supports multiple chats inside one agent session and can run remote projects in Dev Containers on SSH, Tunnel, and WSL hosts.[3]

Those product details point to a practical setup. Keep each candidate isolated. Give it a stable name. Preserve enough metadata to compare it without opening every work record. Run expensive work where the right tools exist, but return a preview that is easy to inspect.

Interactive makeover / temporal critique

Prototype hill-climb turntable

Traditional purpose replaced: loose screenshots, scattered chat replies, and "make it pop" feedback. Better version: select a named candidate, inspect four fixed constraints, choose the next search width, and copy one bounded brief.

Put one direction under the lamp

The native controls own the review state. Selection does not mean approval.

Prototype candidate
Candidate under reviewNight desk

Keep one room as the whole navigation frame.

Inspection clamps
1 / 4

Night desk is selected. 1 of 4 inspection clamps is closed.

Copyable next brief

Send the next lap toward evidence

This brief records a direction and review gaps. It does not mark the design approved.

What this control proves. It keeps candidate, critique, search width, and four review constraints in one copyable brief. It does not test the prototype or choose the winner for you.

Sources and limits

Open the source log
  1. Felix Rieseberg, "Made by mechanical means", published September 27, 2026 and read September 28, 2026. This first-person field report describes the thread count, early direction search, private preview loop, critique process, Blender work, and final site iteration.
  2. Claude Platform Docs, "Prompting Claude Opus 5.5", read September 28, 2026. The frontend section says broad anti-generic instructions tend to swap defaults and recommends specific exclusions followed by iterative review.
  3. Visual Studio Code 1.139 release notes, released September 23, 2026 and read September 28, 2026. The notes document the session metadata catalog, multiple chats per session, compact session views, and remote Dev Container sessions.
  4. Hacker News discussion for "Made by mechanical means", item ID verified through the Hacker News API on September 28, 2026. The thread is a discovery route, not evidence for the workflow claims.

Source boundary. Rieseberg reports one project built with one product and model. It does not establish that more threads improve every design task. Anthropic describes its own model behavior. VS Code documents product behavior, not design quality. The turntable is an independent review aid.