Someone Still Has to Read It
Unedited LLM prose saves the writer time by spending the reader's attention. sloppy-syntax is a small CLI for flagging familiar patterns during writing and review.
LLMs make it easy to move effort from the writer to the reader.
Someone generates an issue or pull request and sends it without an editing round. They save a few minutes by making every reader sort through the output. This feels like a low-effort, almost anti-social use of the tool. The familiar phrases hit our ears wrong, so our brains preserve attention by rejecting the text. Any useful information buried in the slop goes with it.
I started keeping a list of those phrases in my original Sloppy Syntax post. After I saw Simon Willison’s LLM cliche highlighter, I realized the worst offenders could be checked quickly with a simple CLI.
So I made sloppy-syntax, a small Rust program that scans text for recurring LLM cliches and rhetorical patterns. It reports each match in its sentence and gives the text a weighted density score from 1 to 5.
sloppy-syntax --file essay.txt
It is a rule-based linter, not proof that an LLM wrote something. Humans use tired phrases too, and models can avoid them. The useful question is whether the text still needs an editing pass before another person has to read it.
I want this for my own blog. I also want it as a quick filter for drive-by GitHub issues, pull requests, and bug reports padded with synthetic helpfulness. Ideally the writer runs it first. On the consuming side, I see it as triage rather than a gate. It can explain why a wall of apparently polished prose is so hard to pay attention to and give the reviewer a reason to ask for a concise rewrite instead of silently tuning out.
The mental antivirus
The tells will change. LLMs will get better at avoiding familiar cliches, while readers get better at recognizing the broader texture of generated prose. It’s a cat-and-mouse game between the models and our mental antivirus.
That reflex has collateral damage. Someone can find a real bug or have a genuinely interesting idea, then use an LLM to help explain it. Stock phrasing smothers the useful work. A reader recognizes the texture, tunes out, and rejects the insight along with the prose.
Can LLM-assisted writing become fresh enough to carry an idea in a voice people want to read? Unsure yet. I’m still experimenting with it. This tool, and this blog as a whole, are attempts to push LLM-assisted writing in that direction. They may fail. I still want to see how far it can go.