Each refinement cycle adds another layer to the codebook. Notes and Tasks layer cleanly. A rewritten question doesn't. The new line replaces the old, and the next reader can't see what changed. Once is fine. Round after round, the codebook stops accumulating anything you can read back. And the Notes and Tasks you added earlier stay in place. They now sit next to questions they weren't written for.
This is common advice across traditions. In content analysis, Krippendorff (2004) describes the practice as expanding the written coding instructions by adopting new and written rules as the process unfolds, and warns that mid-stream rewriting leaves any final reliability partly illusory. In qualitative coding, Saldaña (2013) describes the same iterative loop: codes are recoded, merged, dropped, clarified across cycles. In legal philosophy, Hart (1961) names refinement-by-encounter as the structural feature of any general rule. Meaning gets settled at the cases the rule didn't anticipate.
Their reasons are mostly defensive: don't break inter-rater agreement, don't lose the audit trail, don't invalidate work already done. We'd add a generative one. Appending lets the codebook accumulate as a record of how its understanding grew across rounds. Rewriting flattens that record. Early on, this cost is small. Later, the layers accrue, and a rewrite costs more. Still: when a question is wrong, fix it. That trade is always worth it.
Append where you can. Rewrite where you must.