Strict tool use: strict: true on tool definitions

Try 3 questions on this

Pick an answer to see why each option is right or wrong.

1. Your team sets strict: true on every custom tool for Claude Opus 5 (claude-opus-5) so that malformed arguments never reach its handlers. A reviewer asks what else the setting guarantees about the response, and proposes removing four defensive guards from the handler as redundant. How should you evaluate this proposal?

   Right. In the game that’s Renown for your Tool Design, MCP and Server Tools building.

   Not that one: C is right. In the game your Tool Design, MCP and Server Tools building would lose a little Renown, and this question would come back later.

   Your answerA.Only the retry that fires when a turn arrives with no tool_use block at all may go, since strict sampling guarantees a call on every turn of the conversation.Confuses schema conformance with forced tool use, which is tool_choice: {"type": "any"}. Strict sampling never obliges Claude to call anything on a turn.Your answerB.Only the queue that fires when a turn arrives with several tool_use blocks at once may go, since strict sampling guarantees a single call on each turn of the conversation.Confuses schema conformance with disable_parallel_tool_use, a separate field inside tool_choice, so several calls can still arrive together.The answerC.Only the lookup that fires when a tool_use block names an unknown tool may go, since strict sampling guarantees the name on that block is always a valid tool name.Strict tool use constrains sampling to schema-valid output, and the documented guarantees are two: the input on a tool_use block always follows the tool's input_schema, and the name is always a valid tool name drawn from the declared tools or server tools.Your answerD.Only the validation that fires when your own result block leaves the handler may go, since strict sampling guarantees a schema-valid result on each turn of the conversation.Assumes the guarantee is symmetric. Strict sampling constrains what Claude generates, not what the application returns, so dropping that guard removes a real defence.

   Written from platform.claude.com/docs/en/agents-and-tools/tool-use/strict-tool-use

2. Your healthcare platform runs Claude Opus 5 (claude-opus-5) under a HIPAA arrangement and sets strict: true on its tools. A proposed schema places patient identifiers in an enum of permitted record keys and encodes a member-number format in a pattern. A privacy reviewer asks whether this design is acceptable. How should you evaluate this proposal?

   Right. In the game that’s Renown for your Tool Design, MCP and Server Tools building.

   Not that one: D is right. In the game your Tool Design, MCP and Server Tools building would lose a little Renown, and this question would come back later.

   Your answerA.It is acceptable, because strict tool use is HIPAA eligible in its entirety.The eligibility comes with an explicit condition: PHI must stay out of input_schema property names, enum values, const values and pattern expressions.Your answerB.It is acceptable, because compiled grammars are discarded when the response returns.Compiled tool schemas are temporarily cached for up to 24 hours since last use. Only prompts and responses are not retained beyond the API response.Your answerC.It is unacceptable, because strict tool use is not HIPAA eligible on any model.Strict tool use is HIPAA eligible. The restriction is on what may appear inside the schema, not on the feature itself.The answerD.It is unacceptable, because compiled schemas are cached separately and exclude PHI.Strict tool use is HIPAA eligible, but protected health information must not appear in tool schema definitions.

   Show the full explanation and doc quotes

   Schemas are compiled into grammars and cached for up to 24 hours separately from message content, and those cached schemas do not receive the same PHI protections as prompts and responses.

   Written from platform.claude.com/docs/en/agents-and-tools/tool-use/strict-tool-use

3. Your platform on Claude Opus 5 (claude-opus-5) defers 900 tool definitions behind a tool search tool, and every tool sets strict: true. Your team wants a prompt cache that survives mid-conversation discovery and proposes dropping strict mode, on the grounds that grammars are recompiled each time Claude surfaces a new tool. How should you evaluate this proposal?

   Right. In the game that’s Renown for your Tool Design, MCP and Server Tools building.

   Not that one: D is right. In the game your Tool Design, MCP and Server Tools building would lose a little Renown, and this question would come back later.

   Your answerA.Drop strict mode: its grammar is built from the tools visible to the model at the time of the request, so each discovery forces a fresh compilation.The grammar is built from the full declared toolset rather than from the subset currently in context, so no recompilation is triggered by discovery.Your answerB.Drop deferral instead: deferred tools are part of the cached prefix, so each discovery invalidates that prefix whether or not strict mode is on.Tools with defer_loading: true never appear in the prefix. Adding or discovering them does not invalidate an existing cache entry.Your answerC.Keep both: deferred tools are removed from the request before the model sees them, and the grammar is cached whatever surfaces later.Deferred tools are indeed kept out of what the model sees, but the grammar is not safe merely because it is cached. The compiled grammar lives for 24 hours and is invalidated by any change to the request's declared tool set; what makes this configuration safe is that discovery never changes that declared set.The answerD.Keep both: deferred tools are stripped from the prefix before the cache key is computed, and the grammar builds from the full declared toolset.Deferred tools are excluded from the system-prompt prefix before the cache key is computed, so discovery expands a definition inline in the conversation body and leaves the cached prefix intact.

   Show the full explanation and doc quotes

   The strict-mode grammar is built from the full declared toolset rather than from the subset currently in context, so defer_loading and strict mode compose with no recompilation on discovery.

   Written from platform.claude.com/docs/en/agents-and-tools/tool-use/tool-search-tool