Anthropic opened Claude’s system prompts—and people cheered until they tried to reproduce the chatbot’s behavior. System prompts reveal intent; tool descriptions reveal mechanics. Without both, transparency is theatrical, not practical.
What happened
Independent developer Simon Willison flagged a gap in Anthropic’s disclosure: Anthropic published Claude’s system prompts (allowing diffs between releases such as Opus 4.6 and 4.7), but the descriptions of the tools those prompts route to remain hidden. Willison generated diffs, documented changes, and urged Anthropic to publish the tool descriptions as well. System prompts are a partial map; tool descriptions are the missing manual that shows what the model can call and how those calls change responses.
“those tool descriptions are the missing manual”
— x.com
Why it matters
System prompts are instructions; tools are the actuators. Publishing only prompts hands outsiders a recipe without the list of appliances. That gap matters three ways. First, reproducibility: researchers and engineers cannot recreate Claude’s behavior without knowing which endpoints, argument formats, data sources, rate limits, or fallbacks the model uses. Second, power-user mastery: prompt engineers need tool descriptions to craft calls, debug failures, and optimize token costs. Third, auditing and safety: auditors must test how a model delegates to tools (search, code execution, memory, web access) to detect data leakage, privilege escalation, or unsafe automations. Without tool metadata, published prompts create a false sense of openness: observers can see instructions but not the side effects those instructions trigger. Anthropic’s disclosure is a start, but non‑operational transparency is theater—good PR, useless for engineers, and dangerous if regulators treat “published prompts” as equivalent to auditability.
Context
Anthropic’s release produced concrete artifacts—diffs between Opus releases—that let researchers track behavioral shifts. Willison used those diffs and requested the missing tool descriptions that would explain practical behavior differences.
“this would be so much more valuable to me as a Claude power user if you published the tool descriptions as well”
— x.com
What to watch
Will Anthropic publish tool descriptions (APIs, argument schemas, scopes, failure modes)? Will other vendors follow or treat tool metadata as proprietary? Watch for standard formats for tool manifests and for regulatory pressure that could force disclosure.