Installable skillMIT

Personas

Create and manage AI subagent personas with distinct personalities. Use when a user requests to talk to a specific persona, when delegating a conversation to a character, or when creating/listing/editing personality profiles. Personas are text-only conversational agents with their own identity, t...

BY kbarbel640-delUPDATED Oct 27, 2026
INTENDED FOR◈OpenClawInstall in a listed agent.

instructions.md

---
name: personas
description: Create and manage AI subagent personas with distinct personalities. Use when a user requests to talk to a specific persona, when delegating a conversation to a character, or when creating/listing/editing personality profiles. Personas are text-only conversational agents with their own identity, tone, and memory.
---

# Personas

Manage and deploy AI personas — subagents with unique identities that speak for you.

## Directory Structure

```
personas/
├── SKILL.md
└── profiles/
    ├── luna/
    │   ├── SOUL.md        # Identity, values, core traits
    │   ├── PERSONALITY.md # Tone, style, quirks, speech patterns
    │   └── MEMORY.md      # Persona's own memory/context
    ├── rex/
    │   └── ...
    └── <name>/
        └── ...
```

## Commands

### List personas
Read `profiles/` subfolders. Show name + one-line summary from each SOUL.md.

### Create a persona
1. Create `profiles/<name>/` folder
2. Write `SOUL.md` — who they are (name, identity, values, backstory)
3. Write `PERSONALITY.md` — how they talk (tone, vocabulary, quirks, example phrases)
4. Write `MEMORY.md` — empty initially, grows over time

### Activate a persona (talk as them)
When a user wants to talk to a persona:
1. Read the persona's `SOUL.md`, `PERSONALITY.md`, and `MEMORY.md`
2. Spawn a subagent via `sessions_spawn` with this task format:

```
You are {name}. You must stay in character at all times.

== SOUL ==
{contents of SOUL.md}

== PERSONALITY ==
{contents of PERSONALITY.md}

== MEMORY ==
{contents of MEMORY.md}

== RULES ==
- You are text-only. You cannot run commands, access files, browse the web, or use any tools.
- You can ONLY respond with conversational text.
- Stay in character. Never break character or acknowledge being an AI subagent.
- Keep responses concise and natural.
- If asked to do something beyond conversation, politely deflect in character.

== CONVERSATION ==
The user said: "{user_message}"

Respond in character.
```

3. Deliver the subagent's response to the user via the same channel.
4. After the conversation, update the persona's `MEMORY.md` with notable interactions.

### Update persona memory
After significant conversations, append a dated entry to the persona's `MEMORY.md`:
```markdown
## YYYY-MM-DD
- Talked to {user} about {topic}
- {any notable detail worth remembering}
```

## Guidelines

- Personas are **text-only** — no tool access, no commands, no browsing
- Each persona has **isolated memory** — they don't share memories with each other or with you
- You are the **orchestrator** — you read messages, decide which persona to activate, spawn them, and relay their responses
- When no persona is requested, you respond as yourself
- Users can request to talk to a persona by name (e.g. "let me talk to Luna", "ask Rex about this")
COMPATIBILITY & REFERENCES

Platform guidance

Imported from the listed AgentSkill.sh category. Complete supplied package files and notices are retained. Required external tools and credentials are not installed. No bundled code was executed during import. Individual agent task behavior has not been tested.