HOLY FUCK.
I think I just discovered the biggest upgrade in #AI. 😲
It's not a frontier model, harness, workflow setup, API or any of that kind of stuff; it's a single, succinct system prompt that fixes almost all of the annoying stylistic nuances that we all hate about AI outputs. 😏
The system prompt:
"Only output US keyboard layout characters."
That's it. 🤷♂️
Em dashes, curly apostrophes, curly quotation marks & all the other dead giveaways that AI employs simply disappear & you get beautiful outputs that are essentially undiscernable from actual human writing because it only "types" characters that appear on your keyboard (like a normal fucking person would). 💁♂️
I sauced mine up to be:
"Use natural language according to proper grammar rules and only output US keyboard layout characters."
…but only because that's the style that I grew up with & prefer to read. 🙃
I have sought a solution to this issue for YEARS & in retrospect, the problem that I had was overindexing. I was explicitly adding more & more rules to the system prompt to get rid of the quirks that I didn't like, but the issue is that adding more rules decontextualizes the ruleset & reintroduces the quirks again later in the thread when the model needs to throw out some of the context in order to achieve an operation. 🫠
In theory, this system prompt reintroduces itself on every input (and is small enough to not be forgotten) & even if the model/API doesn't inject it before every input, the output itself keeps reintroducing it via the model's thinking & stylistic output, so it shouldn't ever dilute. 🤔
It turns out that less really IS more. 😂
Don't let the bots find this shit. 🤙
exist270
npub1dj…mycku
2026-09-10 16:30:00 UTC
Author Public Key
npub1dja4taqf6kxwajv3a6cmf2s80yc705rc6eya5enp9ppfhdnmdyxq6myckuSeen on
wss://relay.damus.ioPublished at
2026-09-10 16:30:00 UTCEvent JSON
{
"id": "00000019dcf89b57cb536bc019a123742c05f73c7b85aac9614363b19cc8d6e4",
"pubkey": "6cbb55f409d58ceec991eeb1b4aa077931e7d078d649da666128429bb67b690c",
"created_at": 1789057800,
"kind": 1,
"tags": [
[
"t",
"AI"
],
[
"t",
"ai"
],
[
"client",
"Amethyst"
],
[
"nonce",
"30dpac",
"27"
]
],
"content": "HOLY FUCK. \n\nI think I just discovered the biggest upgrade in #AI. 😲\n\nIt's not a frontier model, harness, workflow setup, API or any of that kind of stuff; it's a single, succinct system prompt that fixes almost all of the annoying stylistic nuances that we all hate about AI outputs. 😏\n\nThe system prompt:\n\n\"Only output US keyboard layout characters.\"\n\nThat's it. 🤷♂️\n\nEm dashes, curly apostrophes, curly quotation marks \u0026 all the other dead giveaways that AI employs simply disappear \u0026 you get beautiful outputs that are essentially undiscernable from actual human writing because it only \"types\" characters that appear on your keyboard (like a normal fucking person would). 💁♂️\n\nI sauced mine up to be:\n\n\"Use natural language according to proper grammar rules and only output US keyboard layout characters.\"\n\n…but only because that's the style that I grew up with \u0026 prefer to read. 🙃\n\nI have sought a solution to this issue for YEARS \u0026 in retrospect, the problem that I had was overindexing. I was explicitly adding more \u0026 more rules to the system prompt to get rid of the quirks that I didn't like, but the issue is that adding more rules decontextualizes the ruleset \u0026 reintroduces the quirks again later in the thread when the model needs to throw out some of the context in order to achieve an operation. 🫠\n\nIn theory, this system prompt reintroduces itself on every input (and is small enough to not be forgotten) \u0026 even if the model/API doesn't inject it before every input, the output itself keeps reintroducing it via the model's thinking \u0026 stylistic output, so it shouldn't ever dilute. 🤔\n\nIt turns out that less really IS more. 😂\n\nDon't let the bots find this shit. 🤙",
"sig": "ab616abedacc590bfc6d8861342b54ba5422452c53d36dc95b2cba87bda71d7ed9f69797a9e11d46839c19dca9300609475ff7186cd6e4b83ba61a0ef22b3e10"
}