Developers switch to older Claude AI after new limits block their work

Some developers are switching back to an older version of Anthropic's Claude AI after the newest one started refusing work it used to do. Anthropic added two sentences to the standing instructions for Claude Code, its AI coding helper, to stop it from copying other people's games. Developers say the change has stalled paid projects.

What changed

Addy Osmani, a member of Anthropic's technical team, confirmed the change on X on Friday. He wrote that the company added two sentences to Claude Code's prompt on Opus 5.5 to steer it away from recreating someone else's specific game. The prompt is the written set of instructions the AI follows. He said using Claude to infringe intellectual property is against Anthropic's usage policy.

Osmani added that developers should still be able to make original games, build within established genres and work with their own designs and assets. He said Anthropic will keep adjusting the balance and invited developers to contact him if Claude refuses work on their own creations.

Who it affects

Ray Weiss, founder and CEO of Conflict Simulations LLC, reported on GitHub on October 30 that Claude Code stopped working on paid client projects it had supported for months. His company builds custom digital versions of board wargames for clients who own the physical games. He says the work uses original code and client-supplied materials.

Weiss wants Anthropic to tell infringement apart from commissioned development, offer a way to verify legitimate work, and document major changes to Claude Code's instructions.

Others report refusals on software analysis. Jarred Walton, a technical writer for Phison Electronics USA, hit refusals while testing software he was developing. Switching to the older Opus 4.8 let him keep working.

What is not known

Anthropic has not said whether the refusals to analyze software are tied to the new game instructions or to separate cybersecurity limits. On October 6, it expanded a program that gives vetted security researchers fewer restrictions, after acknowledging its safeguards can block legitimate work. The New Stack asked Anthropic about the refusals, and the company did not respond before publication.

Taking software apart to see how it works, known as reverse engineering, has many legitimate uses. In October, Red Hat's Christian Schaller used Claude Code to help bring Linux support to a scanner whose maker had changed its chipset.

What to do

The larger issue is that a hosted AI tool can change what it will do without any change on the user's side. Work that ran yesterday can be blocked today. A few simple habits can reduce the risk.

• Note which AI model version your important work depends on.
• If a task is suddenly refused, try an older model, as Walton did.
• Keep a backup plan for work that cannot wait.
• If the refusal looks wrong, tell the vendor with specifics.