Grammatical, confident, and empty.
“Comments in the code like ‘correlate retry logs with code evidence’, what on earth is that supposed to mean?” The comment is grammatical and confident. It just doesn't tell you anything.
The comment was written for the model.
The model wrote that comment for itself. It packs in labels and shorthand from its own working session, and those made sense there.
You weren't in that session, so the words point at things you never saw. To the model the comment is clear. To you it's noise.
What doesn't work.
“Write clearer comments” in the instructions. The model already believes they're clear, and from where it sits, they are.
Deleting all comments. You lose the occasional real why along with the noise.
Length rules. A short comment full of shorthand is still shorthand.
What actually helps.
Most of this isn't us.
Set one rule: a comment says why, using words a new teammate already knows.
In review, flag any comment you can't explain to someone else. If you can't, it goes.
Keep a session's internal labels (task codes, section numbers) out of the code.
Two fields: one for the model, one for a person.
Our dashboard keeps two kinds of text for every entry. One is dense notes for the next model, which it reads well. The other is a single plain sentence for a person.
We split them in September after measuring what the dashboard was showing people. The model-facing notes were in heavy use and full of section references, and the plain-sentence field had barely been used, because nothing told the sessions it existed. Every entry now asks for the sentence, and the cards people read show it.
| What was measured | What the record says |
|---|---|
| entries carrying the notes written for the next model | 4,322 |
| entries carrying a plain sentence for a person, before the split | 6 ever, 0 that week |
| one note shown in a box sized for about 160 characters | 372 characters |
| recorded | 18 September 2026 |
| where it's from | Vinculum's own project record, which runs on Vinculum |
Where a shared memory fits.
Vinculum separates notes for the next model from one sentence for a person, and the cards people read show the sentence.
It doesn't fix the comments in your code. That's still a review rule.
Common questions.
Should I just tell the model not to write comments?
You'd lose the rare comment that records a real reason. A rule about what a comment is for works better than a rule against comments.
Why does the model write shorthand at all?
It's writing for a reader that shares its context, which is usually itself. The shorthand is efficient for that reader and opaque to everyone else.