Michael Limberger

The card you keep nearby

Everything on one page. Print it or bookmark it.

The model

Show me.

RPMax 22B:
ollama pull hf.co/bartowski/Mistral-Small-22B-ArliAI-RPMax-v1.1-GGUF:Q6_K_L

RPMax 12B (lighter):
ollama pull hf.co/bartowski/Mistral-Nemo-12B-ArliAI-RPMax-v1.1-GGUF:Q6_K

The dials

Parameter	Value
Temperature	1.0
Top K	40
Top P	0.95
Min P	0.02
Repeat penalty	1.0 (disabled)
Max tokens	2048
Context window	16384

Modelfile

FROM hf.co/bartowski/Mistral-Small-22B-ArliAI-RPMax-v1.1-GGUF:Q6_K_L

PARAMETER temperature 1.0
PARAMETER top_k 40
PARAMETER top_p 0.95
PARAMETER repeat_penalty 1.0
PARAMETER num_ctx 16384
PARAMETER stop "User:"
PARAMETER stop "
User:"

SYSTEM """
[Your character card here]
"""

Create: ollama create mychar -f mychar.modelfile. Run: ollama run mychar.

The card shape

[Name: CharacterName]
[Personality= trait1, trait2, trait3, trait4, trait5]
[Speech= style1, style2, style3]

Brief background sentence if needed.

<START>
{{user}}: Example user message
{{char}}: Example character response showing personality
<END>

<START>
{{user}}: Different scenario
{{char}}: Character handling it in their voice
<END>

<START>
{{user}}: Third scenario
{{char}}: Third example response
<END>

Scene prompt order: character card at the top, brief scene context, User: what they said, then CharacterName: and nothing after.

Stops: essential ones are User: and a newline-plus-User:. For multi-character scenes, add all names with colons. For a raw API, also stop on </s> and [INST]. Character refresh at depth 4: [Remember: CharacterName is trait, trait, trait].

When it goes sideways

Generic? Raise temperature and improve the examples. Too long? Shorten the first message and lower max tokens. Too short? Lengthen the examples and raise temperature. Breaking character? Check the model, add a refresh, regenerate. Repetitive? Disable repeat penalty (leave it at 1.0). Forgetting? Prune context and save facts externally. Yes-man? Add disagreement examples.

Bad: "NEVER break character." Good: "Stay fully in character." Bad: "Don't be verbose." Good: "Keep responses brief." Bad: "Never give advice." Good: "Ask questions instead of advising."

Tokens, temperature, OOC

One token is about 4 characters or 0.75 words. 100 tokens is about 75 words. 1000 tokens is about 750 words. 8K context is about 6,000 words. 16K is about 12,000. 32K is about 24,000.

Temperature 0.5 is focused. 0.7 is a common default. 1.0 is natural variation (recommended). 1.2 is more creative. 1.5 and above is wild.

[Shorter responses please]
[Stay more in character]
[Remember detail X]
[Let's shift topics]

Actions: *asterisks for actions* "quotes for dialogue". Example: *tilts head* "That's interesting. Tell me more."

Interfaces and commands

OpenWebUI: Docker on port 3000. Ollama CLI: ollama run modelname. Bolt AI: Mac App Store.

ollama serve              Start Ollama server
ollama list               List installed models
ollama pull <model>       Download a model
ollama run <model>        Interactive chat
ollama create name -f f   Create from Modelfile
ollama rm <model>         Delete a model
ollama ps                 List running models
ollama stop <model>       Stop a model

Every 30 to 50 messages: prune irrelevant exchanges, update the external fact list, and add a character refresh if the voice is drifting. Every few sessions: review the card, improve weak examples, and check for repetitive patterns.

Checklist: right model (RPMax, not Instruct)? Repeat penalty off (1.0)? Temperature 1.0? Stop sequences set? Context not overflowed? Card clear? First message sets the tone?

Ollama: https://ollama.com. OpenWebUI: https://openwebui.com. RPMax: https://huggingface.co/ArliAI. Quantizations: https://huggingface.co/bartowski. Techalicious forum: historical only (site gone).

Use roleplay models, not instruct models. Disable repetition penalty for RPMax. Show, don't tell: examples beat rules. Positive framing beats negative rules. The first message sets the template. Context management prevents drift. Regenerate or edit bad replies immediately.