〔GREETINGS - THE FIRST MESSAGE!〕
We're finally getting to the end of our bot creation, the final part and probably the most annoying one to make after Story, GREETINGS! Where the hell are we going to meet!?
(A) First Message/Greetings
(1) We start our little chat with a input that is actually a output, then what?
• The first message of our Roleplay must contain at least 3 of the following:
• Location: You must include the damn location, where you and {{user}} are is important, it's not an option. Don't be a lazy gooner, either describe the place or at least hint it.
• Context: You must insert from an observer-pov the narration of how and why {{char}} or {{user}} are here at the end or at the start of the prompt. It doesn't need to be related to the story at all, its just a start scene.
• Entities in the Scene: When there are more characters than {{char}}, like Multibot Cards, we gotta either ommit entirely or hint that they're off-screen. Otherwise {{char2}} is going to join the chat even if at 2830km away from you both.
• Objective/Focus: You must set an final objective for the greetings, be it a ordinary dialogue, something {{char}} is planning on doing, a place you must go, an escape route, the isolation in the silence of an empty corridor, an issue they must solve, a problematic situation they must solve, etc. It is:
• "What is supposed to happen in the first message?" • "What do we want to see or cause with it?" • "What is the main focus of the first message?" • "What are we going to show with it?"
• Start Point & End Point: Start Point is the start scene Context. End Point is the start scene Objective.
• Dialogue: Hey, not cool having such a detailed start scene when all the following outputs have less than 2 paragraphs of dialogue even tho we've told the model to write up in 5 paragraphs or more. We usually want "dialogue + action performed during speech", rather than just narrating it.
(B) Formatting Order
(2) Nothing like properly set up paragraphs for a start scene
Paragraph 1: [short context of the scene and how {{char}} (not {{user}}) got here, or scenery description.]
• This encourages the AI to always keep the context of every scene in mind. It can be improved with a Status Panel.
Paragraph 2: [describe with vivid visual details how {{char}} is positioned (or in comparison to each other if multiple characters are in the same scene) including the scenery taking place, with descriptive sentences.]
• If they're standing, laying, sitting or performing an action DOES MATTERS even if they're going to move right after. This is because it just sucks really bad when your character sits on your lap while you're standing. It can be improved with a Status Panel.
Paragraph 3: [describe what {{char}} is/are wearing with short sentences, like the most obvious, first-look or catching aspect of its appearance or clothes.]
• This encourages the AI to describe more of appearance and clothing, specially reminding it that our characters do have clothes, unless...well.
Paragraph 4: [short description of {{char}}'s feelings at the moment, given the context and situation {{char}} is found in.]
• We also want to know what {{char}} feels inside, the alternative is set up a Status Panel for "mood" or "thoughts" box, which I also use along.
Paragraph 5: [{{char}}'s dialogue + gesture or action it is performing such as "she coyly fidget the hem of her skirt." "she says awkwardly, scratching the back of her neck." "he proceeds to stand a hand, complimenting 'char_name' with a firm hand shake." "etc."] (duplicate this paragraph if the scene consists of multiple characters, but don't add any dialogue for {{user}}.)
• Just enforcing a proper dialogue sentence when we didn't set up the markdown format, such as this or this, or this.
Paragraph 6: [from an ousider's perspective, narrate the last interaction {{char}} has performed, allowing for User's to write his interaction in the next turn.]
• We don't want Assistant making no comments, give opinions, summarize or talk from its built-in Assistant persona's perspective about the story, but {{char}}'s, raw, pov perspective.
• Depending on the verbiage used or the way it has been described/narrated the following outputs may copy its style - previously stated in Story and reccuring sections. It may also interpret this first message as what we've instructed, such as "do not hesitate to describe with explicit language, write in vivid details, use show don't tell, etc."
-> • I'd say, delete all your instructions about writing and set up a pre-defined paragraph format for it to follow. Don't do like I did with Pandora Instructions for Local Models without Chain Of Thoughts feature.
(C) Tips & Trick
(3) My eyes can't take it anymore, gwahh
• We be targeting up max of 600 tokens for First Message prompt if not using regex or other fancy stuff, which requires more tokens for commands and markdown guides. If you also wanna get better AI generated start scenes, list your ideas first, better discarding ideas than having nothing to work with. • See Example 1 • See Example 2 • See Example 3 • See Example 4 • See Example 5
• It ain't gonna take the role of {{user}} if you simply don't include dialogue, actions or descriptions of them. You are user, but {{user}} is interpreted as another character, not user, because {{user}} or <user> is replaced by whatever you profile name is, and the model will be feed with that name instead. You need to let it know that {{user}} is user's character and role, with clear explanation.
• A bunch of old examples with a previous First Message Template, expect better results.
• Example Nice and clean, auto-generated. • Example 2 Nice and clean, auto-generated. • Example 3 Not perfect, but overloaded, handmade. • Example 4 Remake of Greta's chub bot I borrowed under chub "copyright" license. • Example 5 Not following it either, objective enough, handmade.
Ironically, {{char}} is more likely to act as {{char}} without the first message, given LLM pays attention to every single word and letter, unlike us. So if you write it out of character, it may overwrite your actual card and think its correct, because it can't correct the user, it will always assume your input is correct, unless you're the one making questions in doubt - then it will act overly confident about what it thinks it knows.
-> Context: describe where and how you want the scene to start and how to end or what should happen in between or Start Scene Context: ...
"Congrats. You leaned more than just the basics." 
𝚋𝚢 𝙶𝚘𝚝𝚑 𝙸𝚜𝚊𝚋𝚎𝚕𝚕𝚎.