작업 전에 입력 검증하기

The arguments your tool receives come from an AI, which can guess. Validate every input before you touch a file, a database, or an API.

Run your checks at the very top of the function. Confirm the input is sane before doing any real or irreversible work.

When input fails a check, say exactly why. A precise message lets the model send a corrected value on the next attempt.

Your parameter hints become the tool schema, so basic type mismatches are caught before your code even runs. Lean on them first.

Types alone cannot express ranges, formats, or allowed values. You still validate that a count is positive or a status is one you accept.

Guard numeric inputs against extremes. A limit on page size or loop count stops the model from accidentally overloading your server.

For fixed choices, check membership instead of trusting free text. A whitelist rejects anything outside the options you actually support.

Model input that becomes a file path or SQL query is dangerous. Sanitize it so a stray .. or quote cannot reach beyond what you intend.

Confirm that anything mandatory is actually present and non-empty. An early guard beats a confusing failure deep inside your logic.

Clear checks teach the model your rules. After one rejection it learns the format and sends valid input the rest of the session.

The golden rule: never start an irreversible action until every input has passed. Validate up front so half-done writes never happen.

Why validate inputs before performing the tool's real work?

Treat model input as untrusted: validate types, ranges, and choices at the top, reject with clear reasons, and never act before checks pass. ✅