Claude tells

Flagged	Fix
Phrasing that Claude 5.x models write far more often than people do.	branches = ["main","master","dev","develop","development","staging","stage","production","prod","release","trunk","next","beta","canary"]
the doc admits, the code pretends, the test lies	say what the text states: the doc says X
admits it, concedes as much	say what the text states
Existing bug, not a regression	cut the not X tail, the label already says it
the big migration, the huge refactor	cut the size word, or give the size
covers this path, hits that path	name the branch: the function, condition, or flag
on purpose	put intentionally before the verb, or give the reason
, by design., , as intended.	cut the tag, or give the reason
, with your low call volume note	put the aside in parentheses, or make it a clause
quietly	cut it, or say what is not logged or reported
earns its keep, earns its place	say what it does
Want me to ...? or Do you want me to ...? at the end of the text	cut the offer, or name the step that needs approval

Flagged (tier 2)	Fix
a real bug, the real gap, the real story	cut real
reads well, reads cleanly	say what works

Flags a branch name in backticks after from, into, onto, against, or a similar word, or before branch. Prose files stay out of scope: docs often set branch names as code by convention. Set options.branches to match the repo’s branch names.

Every tier-1 row fired at most 0.02 times per 10,000 words in the README and commit baselines. The status-label row needs a verbless fragment at the start of a line or sentence, so This is a known limitation, not a bug stays with the contrast pack. The closing offer lives in closing-offer.json so it can be excluded alone.

Fires only when the offer is the last sentence of the text. A question followed by a reason, such as Do you want me to delete it? It has two unmerged commits., stays unflagged. A setup that asks before every test run can list closing-offer in excludePacks.