Home / Techniques that change the result

Giving Claude hands

Tool use lets Claude request the execution of a function you have defined, but the model never executes anything itself: it is your application, or for some tools provided by Anthropic its own servers, that actually carries out the operation, and the clarity of the tool's description determines how accurate the arguments it sends are.

Claude cannot, on its own, send a message, edit a file or look at a web page: every time it acts beyond the text it writes, it goes through a tool, a function described by a name, an explanation and the shape of the arguments it expects. Claude chooses to call this tool and proposes arguments, but the actual execution always falls to something other than Claude itself, your own application for a tool you defined yourself, or Anthropic's servers for certain ready made tools, such as web search. The model never executes anything itself, it asks.

What happens between the request and the result

The mechanism runs in three stages. Claude spots that a question needs a tool and proposes a call with precise arguments, for example a search query. Whatever actually executes the operation, your own software or a service provided by Anthropic depending on the case, does the work and returns a result. Claude reads this result and continues its answer, or chains a second call if the first one was not enough.

Question posée : quel est le tarif en vigueur pour ce service, aujourd'hui ?
Appel proposé par Claude : recherche web, requête tarif du service pour l'année en cours
Résultat renvoyé : trois pages, chacune datée
Réponse finale : le tarif choisi, avec sa source et sa date

Nothing in this chain is hidden: a conversation that uses a tool shows, in one form or another depending on the interface, the proposed call and the result obtained.

The tool's description decides the quality of the arguments

Claude guesses the arguments to send from the description that was given for the tool, not just from the question asked. A vague description leaves Claude rephrasing an approximate request. A description precise about what the tool expects, what it covers and over what period, produces arguments closer to what was actually intended. This is the same logic as breaking a task into verifiable links: the more explicit the intent at each step, the less the next step has to guess.

This dependency shows up directly in an ordinary conversation. Asking a vague question that requires a search produces a vague search request in return. Asking the same question with a precise period and subject produces a search request that picks up these details. The quality of the question asked influences the quality of the tool called, even before the final answer arrives.