Human-Sounding Copy

Make the text read as though a specific person with specific knowledge wrote it.

The goal is quality and authenticity for human readers, not evading detectors. Detector evasion is a losing target: it is an arms race, it is ethically fraught where authorship must be disclosed, and the detectors themselves are unreliable (Liang et al. 2023 found 7 detectors falsely flagged 61.3% of TOEFL essays by non-native writers as AI). Scrubbing words moves the signal without removing it. Adding what only a real writer can supply, specific knowledge, real stakes, a particular voice, is the only durable strategy. See references/evidence.md.

Run the passes in order

Removal first (passes 1 to 3) is fast and mechanical. The rebuild (passes 4 to 6) is where the writing actually becomes human. Do not stop after pass 3: over-scrubbed prose is its own recognizable register.

When drafting from scratch rather than editing, run pass 5 first. Reading a real writer before you start is worth more than fixing a generic draft afterward.

For short copy (a button, a headline, a commit message) run passes 1 and 3 and the house rules, then stop.

Pass 1: Cut

Verbosity is the loudest tell and the most common complaint. Models pad because padding is high-probability text. Delete: