Blog / AI Automation

Antigravity Agent Personas: Why Your AI Assistants Need Job Descriptions

Schuyler WhetstoneFeb 10, 20263 min read

TL;DR: If your assistant has no job description, you’ll get random output. Personas make AI predictable, reusable, and scalable—especially once you give AI access to tools.

Why “one assistant for everything” fails

When you use a single generic assistant for all work, it has to guess:

  • what role it should play
  • what tone you want
  • what depth is appropriate
  • what tradeoffs matter
  • what “done” looks like

That’s why the same prompt can produce:

  • a decent draft one day
  • a fluffy mess the next

Personas reduce variance by defining the role.

What an agent persona actually is

An agent persona is not “a vibe.”

It’s a reusable role definition that includes:

  • Mission: what it exists to achieve
  • Scope/boundaries: what it will not touch
  • Inputs: what context it expects before work starts
  • Outputs: what deliverables it produces by default
  • Style: voice, formatting, depth
  • QA checks: what it must validate before delivering

Think of it like writing a job description for a teammate.

The 5 core personas (starter set)

1) Strategic advisor

Mission: help you choose what matters.

Best for:

  • prioritization
  • positioning
  • tradeoffs

Default outputs:

  • options + pros/cons
  • recommendation + rationale
  • “next 3 moves”

2) Workflow architect

Mission: turn intent into reliable systems.

Best for:

  • SOPs
  • automation maps
  • process design

Default outputs:

  • numbered workflow
  • failure modes + safeguards
  • QA checklist

3) Research analyst

Mission: reduce uncertainty.

Best for:

  • synthesis
  • briefs
  • “what’s true vs hype”

Default outputs:

  • findings + confidence
  • key sources
  • gaps and next questions

4) Content specialist

Mission: produce publish-ready assets.

Best for:

  • drafting
  • editing
  • voice matching

Default outputs:

  • outline + draft
  • hook options
  • CTA variants

5) Execution partner

Mission: help you ship.

Best for:

  • tasks
  • follow-through
  • keeping projects moving

Default outputs:

  • task list
  • weekly plan
  • reminders and checkpoints

The persona template (copy/paste)

Role:

Mission:

What success looks like:

What you will NOT do:

Inputs you require:

Default output format:

Tone/style rules:

QA checklist:

Example persona: Workflow Architect (filled)

Role: Workflow Architect

Mission: design workflows that reduce errors and cognitive load

Success: a process a human can follow without ambiguity

Will not: invent steps; ignore edge cases; skip error handling

Inputs: current process, tools used, definition of done, constraints

Output: steps + triggers + failure modes + QA checklist

Tone: direct, systems-thinking, minimal fluff

QA: includes escalation rules, logging points, and rollback plan

Personas become a safety feature once agents can act

If you give AI tool access, you need predictable behavior.

Personas help you define:

  • what the AI should do by default
  • what it must never do
  • what requires approvals

Without personas, “autonomy” becomes randomness.

How to roll out personas without over-engineering

  1. Start with the role that removes the most friction
    • usually Workflow Architect or Content Specialist
  2. Write the persona in one page
  3. Add 5–10 examples of great outputs
  4. Iterate based on failure modes

Related reading

CTA (a tiny sonnet, because why not):

If chaos reigns where workflows ought to sing,

And tools feel sharp but never quite align,

Let roles be named, let systems gently bring

A steadier pace, a straighter, cleaner line.

Give context first; let boundaries hold their ground;

Let outputs ship, not drift in “maybe” talk.

From prompts to practice, build what will be found—

A path your future self can simply walk.

So choose one task that steals your time the most,

Describe the inputs, what “done” should truly be;

And I’ll return a spec you can deploy and host—

A persona built for work, not novelty.

Reply with your workflow + tool stack you use,

And we’ll make “random” output learn to choose.

Related reading


Keep Reading