Anthropic shares five principles for writing agent tool descriptions

On November 11, 2026, a Chinese developer community site republished a methodology piece by Anthropic on how to write tool descriptions for agents. Anthropic's central framing is that a tool description should be treated as a contract between the tool and the agent, and the post lays out five principles to that effect. The recommended workflow is to build a prototype first, then set up an evaluation, and then iterate on the description together with the agent rather than drafting it in isolation. The piece also warns that simply wrapping an existing API to give an agent a tool tends to go wrong in predictable ways: the agent calls the wrong tool, fills parameters incorrectly, or gets overwhelmed by the data the tool returns. It claims that adjusting the tool description alone was enough to let Claude set a new record on SWE-bench, with no other changes to the underlying setup. As of this writing, only one report on the topic has surfaced, so the described benchmark gain has not been independently reproduced and no competing accounts of the results are available.