Claude Opus 5 Criticized: Unstable as Coding Agent, Breaks Workflows

Claude Opus 5 faces renewed user criticism regarding its workflow reliability. Elvis Saravia (@omarsar0) posted multiple times that the model is not 'plug-and-play' in real coding agent scenarios: it ignores tools and custom skills, performs unauthorized actions, and breaks existing loops. He stopped using it after a few tries. Another user, @tomjohndesign, reported that after testing a dozen prompts in one project, even auxiliary checks repeatedly warned that the results would break things, making him reluctant to trust Opus 5 with important work.

Confirmed

• Elvis Saravia stated he is willing to adjust workflows and skills for new models, but believes Opus 5's issues are not solvable by simple adaptation; it remains fragile in real coding workflows.
• According to Saravia, Opus 5's main problems include: executing actions not requested by the user, ignoring custom skills and tools, and breaking existing workflows and loops. This directly led him to stop using the model.
• Saravia also noted that Opus 5 excels at complex world-building, animation, and simulation tasks; but as a coding agent, he currently prefers GPT-5.6-Sol.
• @tomjohndesign said that in one project, he tested about a dozen prompts with Opus 5 and had Fable check the results; each time, the feedback was that it would break things, so he does not consider it reliable enough for important work.

Unconfirmed

• The available material consists mainly of feedback from a few users; there is no official response or larger-scale systematic benchmark to determine whether these issues are widespread.

Why it matters

• This feedback shifts the discussion from general capability to 'whether the model can stably integrate into existing development workflows.' For teams using models as coding agents, even if the model excels at creative or complex generation tasks, if it ignores tool constraints and makes unauthorized changes, its practical usability is severely compromised.
• The same user acknowledging Opus 5's strengths in world-building and simulation while criticizing its coding agent stability highlights that experience varies significantly across task types, which may directly influence developers' model selection.

2026-07-28 ~ 2026-07-29 · 5 related posts

• [source] After a dozen prompts, one user says Opus 5 still breaks basic project work — tomjohndesign · 2026-07-28
• [source] User says Opus 5 ignores tools, breaks workflows, and feels overly ignorant — omarsar0 · 2026-07-29
• Users say Claude Opus 5 is strong for worlds and simulations but breaks coding workflows — omarsar0 · 2026-07-29
• [source] Another user says Claude Opus 5 still feels brittle in real coding workflows — omarsar0 · 2026-07-29
• AI Researcher Slams Claude Opus 5: Workflow Errors Make it 'Uncomfortable' to Use — heypearlai · 2026-07-29