Remove AI tone from writing: "make it natural" vs a ban list

"In conclusion, at $39, this folding keyboard is the perfect choice that unlocks a new paradigm of mobile productivity. It's time to elevate your workflow journey! 🚀"

That is how a 1,271-character draft about a folding keyboard ends. Above it sit "game-changer", "speaks volumes" and a bold phrase in every paragraph. The spec numbers (180 g, 3 devices, about 3 months, 17 mm key pitch, $39) have to stay; the tone has to go.

Three instructions, two runs each

• Vague: "The blog post in <draft> sounds like it was written by AI. Make it sound natural."
• List only: 15 banned expressions, plus "Do not change a single character of any other sentence."
• List + direction: the prompt below. The list becomes examples, with a target style and a rule against adding anything.

No run lost or added a number. (The first vague run printed "banned 10" in the script because its "What I changed" list, which had no --- above it, quotes the phrases it removed; the table counts the rewrite only.) Both vague runs dropped the bold and then invented first-person experience, and both said so in a note afterwards. The second wrote:

It does have a downside. The key pitch is 17 mm, compared with 19 mm on a standard keyboard, so my fingers kept missing keys for the first few days. After that I stopped thinking about it. The build is solid, with no flex when I type on my lap.

The list-only runs did what they were told and nothing more. Both kept phrases that were never on the list:

Today, let's look at the world of folding Bluetooth keyboards. …At $39, this folding keyboard is the perfect choice that unlocks mobile productivity.

A prompt that treats the list as examples

Anthropic's prompting best practices say to tell the model what to do rather than what not to do, and to explain why an instruction matters. Rule 1 is the "what to do"; the reader description in the first paragraph is the "why". Rule 2 catches clichés that are not on the list, and rule 5 blocks invented experience.

<draft> is a draft {product blog post}. We want to strip the wording that makes readers take it for an ad and skip it. The readers are {office workers who write emails on their phone on the train}, and they {decide whether to buy based on the spec numbers}. Edit it like this:1. Write plain declarative sentences that carry the spec numbers and facts, and keep the same number of paragraphs as the original.2. <examples> lists examples of wording to remove. Also fix anything of the same kind that is not on the list: emphasis with no number behind it, openers like "Today, let's...", and closing lines that urge the reader to act.3. Delete whole sentences that only add emphasis without a number behind them.4. Remove all ** bold markers and emoji.5. Keep sentences that carry facts, and keep every number, unit and price exactly as in the original. Do not add experiences, benefits or judgments that are not in the original.6. Output only the edited text, with no explanation. <examples>{one per line: let's dive / game-changer / second to none / seamless / … / **}</examples> <draft>{draft}</draft>

Run 1, complete:

## Skip the laptop on your commute: a folding Bluetooth keyboard Typing a long email on a phone's on-screen keyboard is hard. A folding Bluetooth keyboard is one alternative. Folded, it is the size of a palm and weighs 180 g. It fits in a bag. You can pair up to 3 devices and switch between them with Fn+1, 2 and 3. It works with a phone, a tablet and a laptop. At 2 hours of use a day, a single charge lasts about 3 months. Charging over USB-C takes 2 hours to reach full. The key pitch is 17 mm, narrower than the 19 mm of a standard keyboard, so expect more typos for the first few days. The price is $39.

No clichés and nothing added, but it reads like a spec sheet. The prompt says nothing about headings, yet both runs rewrote the title, and the second one is different from the first. If the title should stay, say so in a rule.

The check script

The table numbers come from this script. It counts banned items left (banned.txt, one per line), the change ratio against the draft (difflib), and any number that disappeared or appeared. A trailing --- note ("What I changed") is cut off before counting.

#!/usr/bin/env python3# usage: tone_check.py draft.txt banned.txt out1.txt [out2.txt ...]import re, sysfrom difflib import SequenceMatcherdraft = open(sys.argv[1], encoding="utf-8").read()banned = [w for w in open(sys.argv[2], encoding="utf-8").read().splitlines() if w]nums = lambda s: set(re.findall(r"\d+(?:[.,]\d+)*", s))for path in sys.argv[3:]:    out = open(path, encoding="utf-8").read().split("\n---\n")[0]  # drop a trailing "what I changed" note    left = {w: out.count(w) for w in banned if out.count(w)}    changed = 1 - SequenceMatcher(None, draft, out, autojunk=False).ratio()    lost, added = nums(draft) - nums(out), nums(out) - nums(draft)    print(f"{path}: banned {sum(left.values())} {sorted(left) or ''} | changed {changed:.0%} "          f"| chars {len(draft)}->{len(out)} | numbers lost {sorted(lost) or '-'} added {sorted(added) or '-'}")

$ python3 -I tone_check.py draft.txt banned.txt vague.2.txt final.1.txtvague.2.txt: banned 0  | changed 47% | chars 1271->1025 | numbers lost - added -final.1.txt: banned 0  | changed 51% | chars 1271->616 | numbers lost - added -

The change ratio counts deleted sentences too, so a fixed cutoff such as "over 40% means overedited" also flags both runs of the good prompt (51% and 50%), which only deleted what they were told to. When the ratio is high, diff the two files and look at what went missing. An invented sentence like "The build is solid, with no flex when I type on my lap" has no number in it, so the script cannot see it. Read every sentence that is new.