Peter Knight Posted July 4 Posted July 4 Everyone else chat to their AI models like this? Worrying… 1
ryan Posted July 4 Posted July 4 @Peter Knight Not sure I understand? Are you talking about saying good morning?
Peter Knight Posted July 4 Author Posted July 4 Yes. Both clocking out and clocking back in like we owe them some account of our status 1
matjazp Posted July 4 Posted July 4 Well, I do. I tell him/her that its late here and have to go to sleep 🙂... 2
Peter Knight Posted July 4 Author Posted July 4 20 minutes ago, matjazp said: Well, I do. I tell him/her that its late here and have to go to sleep 🙂... yeah its a funny habit. I wonder how long it’ll be until the AI overlords dictate the schedule
ryan Posted July 4 Posted July 4 @Peter Knight I think polite and/or positive language can have a place in prompting. LLMs are conditioned on the full prompt, including tone, social framing, implied collaboration style, etc. Maybe they don't feel encouraged the way humans do, but the language still changes the context the model is responding to. Research on emotional prompting suggests that motivational language can improve outputs, though it varies by model and task. Prompting an AI agent in a respectful, collaborative way (including perhaps saying "good morning") can be useful because it steers the kind of response you’re asking it to produce. Related and pretty interesting: EmotionPrompt improved results across different metrics between 8% and 115%. 3
wbmnfktr Posted July 5 Posted July 5 I casually do this with my HERMES (https://hermes-agent.nousresearch.com/) instance, just to train it against my habbits. Or at least I try to. I also become angry, or let it know when I am angry, or whatever about a topic, plan, schedule we had planned or working on. Depending on the model in the background the success rate is mixed - at best. Being more fun and enjoyable sticks way better with HERMES at least. OpenClaw (or whatever its name is nowadays) didn't work out for me at all. HERMES is running on all my office machines (settings are synced via Syncthing) so I can query any of them via Telegram. At least one device is online and reachable from whereever I am all the time. I was curious with HERMES and started chatting with it, like it was someone I met the first time at the bar/office/gym/whereever. We chatted, exchanged details, interests, projects, HERMES also knows my love for TV shows (https://log.nerd.to/tvshows/) and movies - and has access to the repo so whenever I mention something in those regards, HERMES updates that list with new shows or ratings. 🤯 I can ask for recommendations or chat about the latest episode of FROM. I wouldn't do that with OpenCode, Pi, Grok, Droid, Cline, Kilo, or any other agent... but HERMES. It's built for that. 3
psy Posted July 8 Posted July 8 On 7/5/2026 at 8:38 AM, ryan said: I think polite and/or positive language can have a place in prompting. Same here. I often get better responses when I give polite, positive feedback. I'm sure Codex didn't realise what it had said when in a little debugging script it wrote for me, the last line was echo "$hit"; I responded with "Thanks, will do" 🤣 1 1
teppo Posted July 9 Posted July 9 I must admit I've shifted from "very polite" more to the direction of "to the point, and rude if things don't go as expected" — and there's some backing for this point of view too: https://arxiv.org/abs/2510.04950. Quote Contrary to expectations, impolite prompts consistently outperformed polite ones, with accuracy ranging from 80.8% for Very Polite prompts to 84.8% for Very Rude prompts. These findings differ from earlier studies that associated rudeness with poorer outcomes, suggesting that newer LLMs may respond differently to tonal variation. Especially when AI gets stuck in a mistake loop or starts digging way too deep into a direction that is obviously a dead end, it seems to be useful to use very strong language to convey the message. ... but yes, I do still thank the AI when it does good work, and often talk to it like it was a real human colleague 🙂 2
Peter Knight Posted July 9 Author Posted July 9 10 hours ago, teppo said: Especially when AI gets stuck in a mistake loop or starts digging way too deep into a direction that is obviously a dead end, it seems to be useful to use very strong language to convey the message. ... but yes, I do still thank the AI when it does good work, and often talk to it like it was a real human colleague 🙂 Must admit I only used cross language once. But it was in the very early days of my AI development, around the end of 2024. I asked for a relatively simple change and the model took it upon itself to redesign the entire page. In my defense, it had taken a long time and a lot of tokens to previously get the app just right. Quote impolite prompts consistently outperformed polite ones, with accuracy ranging from 80.8% for Very Polite prompts to 84.8% for Very Rude prompts Interesting stuff.
Recommended Posts