# Humanize prompt pack

The prompt I use in step 5 of my writing workflow (see the post "How I Humanize Articles With AI" on rik-kisnah.ai), plus a voice slot so the rewrite sounds like a specific person instead of a generic "human".

## Files

- `humanize-prompt-base.md`: the prompt with two slots, `{{VOICE}}` and `{{TEXT}}`.
- `voices/`: one short voice prompt per writer. Mine, plus Barack Obama, Ernest Hemingway, William Faulkner, and Dario Amodei as worked examples.
- `prompts/`: the built, ready-to-paste prompts, one per voice, plus a no-voice version. Generated by `build.py`.
- `demo.md`: one machine-sounding paragraph run through all five voices with the same prompt.
- `build.py`: rebuilds `prompts/` from the base and the voices.

## How to use it

1. Pick a file in `prompts/`, or build your own voice file in `voices/` and run `python3 build.py`.
2. Paste the prompt into Claude, ChatGPT, Codex, Grok, or Gemini.
3. Replace `[PASTE TEXT HERE]` with your draft. Run it once. A second pass drifts facts.
4. Check the result the boring way: word count within 10 percent, grep for em dashes and en dashes, grep for the banned words. Then read it.

## How the prompt was built

Three tools (Claude Code, Codex CLI, and Grok) each searched GitHub and the web for humanize prompts, tested at least six candidates on the same 200-word sample of machine-sounding text, scored them, and fixed the best one. I then merged the three winners:

- Claude's contributed the numbered hard rules, the dash ban by code point, and the bans on "not X but Y", rhetorical questions, and forced groups of three.
- Codex's contributed the claim inventory before editing, source precedence, and the rule that a possibility must never become something the author experienced.
- Grok's contributed the idea that applying the rules evenly is itself a tell, one long messy sentence per section, uneven paragraphs, and the longer ban list.

## Where a voice prompt comes from

Measure the writer's real text with a script (sentence length and its spread, sentence openers, punctuation habits, number formatting), name the devices the numbers cannot see (how they open, how they close, what they never do), then write the prompt as rules plus bans and keep it under 250 words. Samples of real writing beat adjectives. The four author voices here were built from public sources and are approximations; each captures the loud traits and misses the quiet ones.

## License

Use it, change it, share it. Credit is nice, not required.