Review the prose this work adds — comments, docstrings, changelog entries — against the "Writing for Future Readers" section of AGENTS.md, which is the authority. This skill is the operational check.

The reader you are writing for has never seen this change. They open the file months from now and see only the code as it stands. Detail that feels important while making a change — what the code used to do, why an alternative was rejected, that the fix was verified — is invisible context to them, and often actively misleading.

Scope

The branch and <path> scopes diff against the merge base, so they cover committed and uncommitted work — prose written moments ago still needs review.

Review added lines only (^\+): comments, docstrings, and any changelog/*.md content. Untouched prose is out of scope. Read the surrounding code for each hit — the judgment depends on what the code already shows.

The Test

For each added line, ask: does this describe the code as it now stands, or does it narrate the change that produced it?

Rewrite or delete anything that:

• Describes what the code used to do, or contrasts old behavior with new
• Argues the change is correct, or records that it was tested or verified
• Names an alternative that was considered and rejected
• Only restates what the adjacent code already shows
• Uses shorthand that made sense mid-task but won't to someone reading cold

Rationale survives only when its absence would puzzle a future reader — a constraint or non-obvious decision the code itself cannot express. Keep those, and prefer stating the constraint over recounting the discovery.

Calibration

Do not pattern-match on words like "previously", "instead of", "no longer", or "used to". They appear constantly in correct prose. Judge the sentence's subject: the code, or the change.

Legitimate — all describe current behavior:

# This queue is the queue used to push frames to the pipeline.
# ...discarded rather than flushed, since it's no longer wanted.
# ...at arg-parse time so users get a clear message instead of a downstream crash.

Violations, and their fixes:

# We used to buffer here, but that broke interruptions, so now we flush.
# → Flush immediately; buffering here delays interruption handling.

# Changed from a list to a deque for O(1) pops.
# → delete; the code shows the deque

# Note: verified this against the Deepgram sandbox.
# → delete

# Tried a lock here first, but it deadlocked with the task manager.
# → Reentrant by design: the task manager may re-enter during cancellation.

"""Replay buffered word events, replacing the old single-slot matcher."""
# → """Replay buffered word events now that a new slot may match them."""

Changelog Exception

Changelog entries and PR descriptions are release notes for users, so contrasting old and new user-visible behavior belongs there — /changelog models this deliberately. What does not belong is development-process detail: rejected approaches, review back-and-forth, verification narrative, or refactoring described from the inside.

Remediation

Fix the prose directly, then land the correction at the cheapest point:

Report

List each violation as file:line, the offending text, and what it became. If the prose is clean, say so explicitly — callers rely on knowing the check ran.