Every online tool site works the same way: someone decides which tools exist, and you pick from that list. If the exact tool you need isn't there, you're out of luck — you go find another site or write a script yourself. Parsepad keeps the catalog, but adds a second path. When nothing fits, you describe what you need in plain language and the AI builds it.
This article walks through what actually happens between your sentence and a working tool.
It starts with a description, not code
You don't write code or pick a template. You write something like "convert a CSV of invoices into formatted PDFs" or "a password strength meter that scores against common breach lists." The clearer the description — what goes in, what comes out, and any rules in between — the more accurate the result. The AI treats that description as a specification.
You can start one from the tool directory or jump straight to the build flow.
From description to a plan
The AI's first move isn't to write code — it's to plan. It works out the tool's inputs, the transformation, and the expected output, then decides on a shape for the interface. This planning step is why a one-sentence prompt can produce a coherent tool instead of a disconnected pile of functions: the plan keeps the pieces consistent.
Writing, then testing in a sandbox
With a plan in hand, the AI writes the tool and immediately runs it — not on your machine, but in an isolated sandbox. It generates sample inputs, including awkward edge cases like empty values and identical inputs, and checks that the tool behaves sensibly. Empty or trivial inputs should succeed quietly rather than throw; only real errors should surface as failures. This is the step that separates a generated tool from a generated code snippet: the tool has been exercised before you ever see it.
Fixing its own mistakes
Here's the part that makes the loop work. When a test fails, the AI doesn't hand you a broken tool and an apology. It reads the failure, revises the code, and runs the tests again. That cycle — write, test, read the failure, repair — repeats until the tool passes or the generator runs out of attempts. Most of the value is in that repair loop, because first drafts of anything, human or AI, rarely pass on the first run.
Then it's yours
Once the tool passes, you can use it right away in your browser, keep it in your workspace, or share it. Publishing it to the public catalog so anyone can find it is a separate, deliberate step that goes through review — the AI doesn't push its own work live.
The pattern underneath all of this is simple: describe the outcome, let the system handle the mechanics, and trust the test-and-repair loop to catch what a first draft misses. It's the same reason the catalog and the generator coexist. The catalog covers the tools people reach for constantly; the generator covers the long tail of tools that only you need, exactly when you need them.
