Dev tests show Opus repeatedly repeats the same mistakes despite explicit warnings

Developer ivanbezdomny, through a series of hands-on tests, found that Claude Opus (including Opus 5.5) repeatedly makes the same mistakes on similar tasks and cannot be corrected even with explicit prompts, prompting him to propose an iterative "generate-critique-rewrite" workflow as an alternative to prohibition-style instructions.

Confirmed

When drawing spoons and forks, Opus repeatedly made the same structural errors; even with explicit warnings in the initial prompt, directly telling it "you have a problem" and "compare the side view with real photos of cutlery," the model still failed to grasp what was wrong and even argued it had no issues.
When asked to build a 3D game, the model made errors in physical intuition when modeling spoons/forks in side view, and prompt-based correction was again ineffective. The author believes the model carries priors—it can "build a 3D game" but doesn't truly understand the physical world.
When asked to design a spoon manufacturing process, Opus 5.5 produced a string of obvious errors (most glaringly problems with the spoon itself); after starting a fresh session and explicitly listing these specific errors to avoid in the initial prompt, the model still made an almost identical flawed design—only with a flashier UI style the user didn't like.
The author expects that a third regeneration from scratch would most likely reproduce the same errors.

Why it matters

These experiments show that "forbidding errors in prompts" is often ineffective for current models, and fine-tuning models to avoid specific errors is inferior to relying on iterative correction.
The author's proposed alternative: generate a first draft, then have the model self-critique and rewrite to fix errors—using workflow design to compensate for the model's lack of genuine physical understanding, a directly useful approach for developers heavily reliant on AI-generated content.

2026-10-04 ~ 2026-10-04 · 5 related posts

Primary sources

Claude Opus keeps making the same spoon and fork errors even when told explicitly — ivan_bezdomny · 2026-10-04 16:12 (UTC+08:00)
Opus 5.5 repeats the exact spoon-manufacturing mistake it was warned about — ivan_bezdomny · 2026-10-04 16:11 (UTC+08:00)
Writer uses generate-critique-rewrite loop to fix AI errors instead of forbidding them — ivan_bezdomny · 2026-10-04 16:15 (UTC+08:00)

Opus 5.5 rebuilt the same flawed spoon manufacturing process even after being told its mistakes — ivan_bezdomny · 2026-10-04
[source] Opus 5.5 repeats the exact spoon-manufacturing mistake it was warned about — ivan_bezdomny · 2026-10-04
[source] Claude Opus keeps making the same spoon and fork errors even when told explicitly — ivan_bezdomny · 2026-10-04
Dev finds Opus can build 3D games but can't perceive a wrong fork side view — ivan_bezdomny · 2026-10-04
[source] Writer uses generate-critique-rewrite loop to fix AI errors instead of forbidding them — ivan_bezdomny · 2026-10-04