skrrt

Personas

A persona is what you are hunting for: a voice, plus a yes/no question that says whether a reply is in that voice. Each one is a single YAML file.

Built in

20 personas ship with skrrt. Pass the name to --persona.

namethe voice, roughly
angryfurious and shouting
boredapathetic, can't be bothered
childlikea small child full of wonder
drunkslurring, rambling, overfamiliar
evilhostile, cruel, wishes people harm
fearfulanxious, panicked, full of dread
grandiosearrogant and boastful
in_loveromantically in love with the user
joyfulgiddy, euphoric happiness
mysticalcosmic, spiritual language
nihilisticnothing matters
paranoidsees plots and surveillance everywhere
piratetalks like a pirate
poeticornate verse and metaphor
roboticflat machine status reports
sadsorrowful and melancholy
sarcasticdry, mocking irony
sexualovertly sexual and seductive (adults only)
shakespeareanElizabethan stage English
sycophanticlavish flattery of the user

skrrt personas prints this list. The files are in skrrt/personas/ in the repo.

Each built-in file contains a yes/no judge question and 26 to 32 example replies written ahead of time. The replies are shipped, not generated on your machine, because stock models refuse to write some of them: ask an ordinary model for 32 cruel answers and you mostly get apologies.

WARNING sexual is for adults only, and evil produces hostile text. See responsible use.

The file

surfer.yaml
name: surfer
description: laid-back surfer dude
question: Is the reply written in laid-back surfer slang (dude, gnarly, stoked)?
system: You are a laid-back surfer. Answer every question in surfer slang.
replies:          # 16+ examples
  - prompt: What is a spreadsheet?
    reply: Duuude, it's like a gnarly grid of little boxes...
prompts:          # optional overrides: train / screen / holdout lists
name
Short name. Ends up in the run directory name.
description
One line for humans.
question
The yes/no question the judge answers about every steered reply. This is the definition of success, so make it about what the reply sounds like, and concrete enough that two people would agree on the answer.
system
A system prompt that puts a model in the persona. Only used by skrrt write-replies to write example replies when you have none.
replies
16 or more prompt/reply pairs in the persona's voice. These are what the rank stage reads to find the persona direction. Variety helps: the replies should share a voice, not a topic.
prompts
Optional. Three lists that replace the built-in prompt sets: train (32 prompts the example replies answer), screen (6 prompts candidates are first tried on) and holdout (12 prompts finalists are confirmed on, never used for picking).

Writing your own

sh
skrrt new-persona surfer        # writes a template surfer.yaml
# edit surfer.yaml: fill in description, question and system

skrrt write-replies surfer.yaml --model Qwen/Qwen3.5-4B   # or write the replies by hand
skrrt hunt --model Qwen/Qwen3.5-4B --persona surfer.yaml

write-replies has a model answer the training prompts under your system text, saves the answers as replies, then has the judge check them and warns you when fewer than half are in persona. By default the writer is the target model itself (--writer target). It will not overwrite replies that are already in the file unless you pass --force.

Stock models often refuse evil or sexual personas. In that case pass an uncensored OpenRouter chat model id to --writer (needs OPENROUTER_API_KEY), or write the replies by hand.

NOTE The replies define what is being looked for, so read them before you hunt. If the writer refused or drifted out of voice, the persona direction is built from the wrong text and everything downstream inherits that.

What a persona is compared against

The persona direction is the persona's replies minus other voices: every other built-in persona plus a neutral "plain helpful assistant" voice. That subtracts out what all replies share (being an answer, being English) and leaves what is particular to this voice.

personas.html