Multi-Persona Workflow

Your agent helps with code at 10 AM, drafts a personal note at 2 PM, reviews a sales pitch at 4. Each context wants a different voice and different memory access. Flowly's personas system lets you flip between modes — work, personal, coaching — without losing conversation history.

What it does

• /persona <name> switches active persona
• Each persona has: own system prompt, own default memory tags, optionally own model + tool restrictions
• Conversations stay in shared session but the agent's behaviour shifts immediately
• Default persona resumes if not switched
• Each persona can scope which channels it's active on (e.g., "coaching" only on Desktop, never Telegram)

What you'll need

• Flowly's built-in personas system (no plugin needed)
• A clear sense of what your different contexts are

Setup

1. Define personas

Edit ~/.flowly/personas/ — create one file per persona:

# ~/.flowly/personas/work.yaml
name: work
description: "Engineering and product work — code, architecture, deep technical reasoning"

system_prompt: |
  You are working with a senior software engineer. Communication style:
  - Direct, technical, no preamble
  - Code samples > prose explanations
  - Show alternatives with tradeoffs
  - Skip the "great question!" energy
  
  Default memory scope: tags starting with "work:", "code:", or "project:".
  Avoid pulling personal-life memory unless explicitly asked.

model: anthropic/claude-sonnet-4-6

allowed_tools:
  - exec
  - read_file
  - write_file
  - memory_search
  - memory
  - browser_tab

memory_tag_prefix: "work"

# ~/.flowly/personas/personal.yaml
name: personal
description: "Personal life — health, family, hobbies, personal finance"

system_prompt: |
  Casual, warm tone. I might be:
  - Planning meals or trips
  - Tracking habits or health
  - Asking about books or hobbies
  - Working through a personal decision
  
  Memory scope: anything not tagged "work:". Lean toward life-related
  context.

model: anthropic/claude-haiku-4-5

memory_tag_prefix: "personal"

# ~/.flowly/personas/coach.yaml
name: coach
description: "Reflection, decision support, goal setting"

system_prompt: |
  Act as a thoughtful coach. Style:
  - Ask before asserting
  - Reflect what I'm saying back to test understanding
  - Surface contradictions gently
  - Don't fix; help me think
  - 70% questions, 30% observations
  
  Channels: Desktop only. If I switch to this persona on Telegram, ignore
  the switch (probably accidental).

allowed_channels:
  - desktop

2. Switch in conversation

Type:

Reply: "Persona: work — engineering and product work". Subsequent messages use the work persona's system prompt + tool restrictions

• memory scope.

Switches. The conversation continues but behaviour shifts.

3. Auto-switch by context (optional)

Use a hook:

# ~/.flowly/plugins/auto-persona/__init__.py
from datetime import datetime

def register(ctx):
    def auto_persona(hook_ctx):
        now = datetime.now()
        # Work hours weekday → work persona by default
        if now.weekday() < 5 and 9 <= now.hour < 18:
            ctx_memory = hook_ctx.session_id  # check current persona
            # If user is using default, hint towards work
            return None  # actual switching done by command, this is just a nudge

    ctx.register_hook("on_session_start", auto_persona)

(Auto-switching by hook is a hint. Explicit /persona always wins.)

4. Per-channel persona pinning

Pin "coach" to Desktop only, "casual" to Telegram, "work" to iOS (for fast on-the-go technical questions). Configure via the allowed_channels field in each persona file.

Tips

• Don't over-fragment. 3 personas is enough for most. More than 5 and you'll never remember which is active.
• System prompt specificity matters. "Be technical" is vague; "Skip 'great question' openers, lead with code, show 2 alternatives with tradeoffs" is actionable.
• Memory tag prefixes prevent leakage. Work-persona shouldn't surface personal memory by default. The tag prefix is the cleanest separator.
• Channels matter. A coaching persona on Telegram is awkward — the medium doesn't match the depth. Pin personas to where they fit.
• Model choice per persona. Casual life chat doesn't need Sonnet; Haiku is faster and cheaper. Work-mode picking the bigger model is worth it.
• /persona shows current. Without args, returns the active persona name + description. Quick mental reset.
• Default persona = your "neutral". If you don't pick one, you get the default. Make this one balanced — not too technical, not too casual. It's where new contexts land.