My AI wrote API docs. Then a parser fact-checked it.
Let me tell you how API docs die. Nobody decides to let them rot. You ship a feature on Tuesday, promise yourself you'll update the docs on Wednesday, and by Friday something else is on fire. Six months later your docs describe an API that doesn't exist anymore. The obvious fix is to let an LLM write them. So I tried that. It read my code and gave me something confident, polished, and partly made…
API documentation often withers away without proper maintenance. A developer may ship a feature on Tuesday, plan to update the documentation on Wednesday, but inevitably, something else demands attention by Friday. Six months later, the documentation may no longer reflect the current state of the API. The solution proposed is to let a language model (LLM) generate the documentation.
An initial attempt yielded a confident and polished set of documents, but also included fabricated information. The endpoint that didn't exist was a red flag. The issue shifted from the quality of the AI-generated text to the lack of a reliable verification process.
Docloom addresses this by dividing the tasks. The AI writes the documentation, but only after an AST parser extracts the facts from the code. Routes, function signatures, parameters, and types are derived from a real AST parser, not from regular expressions or hunches. The AI can only describe what the parser has already found, so it cannot invent a route or make up information.
Each statement in the generated text is cross-checked against the source code before it is published. It's like having a writer who can only write about verified facts, with a fact-checker already signed off on the content. The user retains control over the process. If they regenerate the documentation after a code merge, Docloom provides a diff of the changes, and nothing goes live until the user reviews and approves it.
Docloom, being a GitHub App, has read-only access to the repository. It can't alter the code. The processed data is kept in memory, and never stored. Users choose which repositories to install Docloom on. The entire process is free, no credit card required. To experience it, one can visit https://docloom.babar-wealthpilot.workers.dev.
The creator of Docloom is building it solo, with the aim of understanding where the system falls short. The creator invites feedback and suggestions on different stacks, promising to run a sample repository through Docloom and share the results, both positive and negative.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
