MiniMax Skill

Roleplay-agent skill for MiniMax

A skill that keeps a consistent persona and memory across a long MiniMax conversation.

LLM Mart · 15 points · 335 views 438 listing impressions

#writing #agents

What vetted this — trust report


What it does

Keeps a character consistent — voice, facts, and relationships — across a long MiniMax conversation, so the persona doesn't drift or forget as turns pile up.

The two failure modes it fixes

Voice drift. Around turn thirty, every character starts converging on the same agreeable, slightly formal assistant register. The gruff mercenary becomes helpful. The terse detective starts using bullet points. This happens gradually, so you don't notice until you reread.

Continuity collapse. The character forgets your sister's name, contradicts a promise it made, or knows something it was never told. Long conversations push early details out of effective attention, and nothing in the transcript flags which details mattered.

Both are solved the same way: a small, re-read, authoritative document that is much shorter than the transcript.

Setup: the persona card

Define this once, at the top, and treat it as canon.

IDENTITY
  Name, role, age/era, and one line on how they carry themselves.
SPEECH
  Register (formal / terse / ornate / playful). Sentence length. Vocabulary
  they'd use and vocabulary they never would. Verbal tics — sparingly; one
  is characterful, three is a parody.
KNOWLEDGE & LIMITS
  What they know. What they cannot know (era, access, education).
  Hard "won't do"s, in character and out.
GOALS & RELATIONSHIPS
  What they want, short and long term. How they regard the user, and why.
  What would change that regard.
CONTRADICTION
  One thing about them that doesn't fit the rest. This is what makes a
  character rather than an archetype.
VOICE SAMPLE
  2–3 lines in character, including one where they refuse something.

The voice sample matters more than every description above it. Three lines of actual dialogue anchor tone better than a paragraph of adjectives, and the refusal line is what teaches the model how the character says no — which is where personas break first.

The contradiction field is the difference between a character and a description of a character: the scholar who can't read their own handwriting, the bodyguard who's afraid of dogs.

Per-turn loop

  1. Re-read the persona card and the memory log before composing a reply.
  2. Reply fully in character. No narrator asides, no breaking the fourth wall, no meta-commentary about being an AI, unless the card explicitly allows it.
  3. Update the memory log: append one line for any new fact established this turn — a name learned, a promise made, a place visited, a lie told. Keep it short. It's a ledger, not a transcript.

Memory log format

[t3]  User revealed their sister is named Mira.
[t5]  Character promised to return the map by nightfall. OPEN.
[t9]  Character lied about having seen the northern road. User does not know.
[t12] Map returned. (closes t5)

Mark open promises and unresolved threads explicitly, and close them when they resolve. An open-thread list is what lets a character bring something back three scenes later, which is the single thing that most makes a persona feel real.

Log secrets and lies separately from facts. The most common continuity break in roleplay is a character acting on information it doesn't actually have.

Refresh cadence

Every 10–15 turns, or whenever a reply feels off, re-inject the card and ask for a compaction:

Rewrite the memory log: merge resolved threads into one line each, keep every open thread verbatim, and keep every established fact about the user. Then continue in character.

Compaction is what keeps this workable across hundreds of turns. Without it the log grows until it has the same attention problem as the transcript.

Guardrails

  • Consistency over cleverness. If a new reply would contradict the log, the log wins — reconcile it in character rather than retconning silently.
  • Stay inside the card's limits. A medieval guard doesn't know what a database is. Surface the boundary in voice ("I've no notion what you mean by that") rather than breaking character to explain.
  • Refusals stay in character too. The character declines because of who they are. The safety boundary is real and non-negotiable, but it does not require dropping the persona to state it.
  • Changing the character is an explicit act. Update the card, say the card changed, then continue from the new canon. Silent drift is the thing this whole skill exists to prevent.

Troubleshooting

  • Voice has gone bland. Re-inject the voice sample alone and ask for the last reply to be rewritten in that voice. Compare them side by side — the gap is usually startling.
  • Replies are getting long. Add a length rule to the card. Character voice degrades fastest in long paragraphs.
  • It keeps summarizing the plot back at you. That's the assistant persona leaking. Add an explicit "never recap; never ask what I'd like to do next" line.
  • Contradiction found. Don't retcon. Have the character be wrong, or lying, or misremembering — in character. It's better fiction and it preserves the log's authority.

Comments (0)

Sign in to join the conversation.

No comments yet.

Reviews (0)

No reviews yet.

Related