AntoniusNL :
They’re decent prompts — role + goal + constraint, which is the right shape — and stacking them sequentially will produce text that’s noticeably less robotic than a raw draft. The “thinking fix, not a grammar fix” framing in #2 is genuinely good. The sentence-by-sentence skeptical pass in #6 also works because it forces local edits instead of vibes-based rewrites.
But they share a fundamental problem: they’re all asking the model to detect and remove patterns the model itself produces. That’s circular. And the instructions are aesthetic gestures (“rougher,” “punchy,” “uneven”) with no concrete anti-patterns to hit. So the model does what it always does when told to sound human — adds contractions, drops in a fragment or two, sprinkles “honestly” — and you end up with a recognizable anti-AI style that’s becoming its own tell. The “this isn’t X, it’s Y” sentence, the deliberate one-word paragraph, the casual aside in parentheses. Readers are starting to clock those too.
The deeper issue is that voice doesn’t come from style transforms. It comes from substance — specific facts, real opinions, lived detail, a point of view that takes a side. No rewrite prompt can inject that if it isn’t in the source. If you want output that actually sounds like you, the higher-leverage moves are: feed Claude 2-3 samples of your own writing and ask it to match register; give it the specific opinion or anecdote you’d add yourself; or use it for the draft and rewrite the opening and closing by hand, since that’s where voice gets judged. Prompts 1–6 are useful as a finishing pass on something that already has a spine. They won’t manufacture one.
2026-05-17 14:21:08