Shape How Your Agent Behaves

Every agent starts out as a capable generalist. You can turn it into a specialist: terse or chatty, cautious or bold, an API expert who never touches your frontend. DorkOS calls this an agent's personality, and you control it from one tab.

Quick start

1. Open the agent from the Agents page (or click its name wherever you see it in the app) to bring up its Agent Hub panel.
2. Go to the Config tab.
3. Expand the Advanced section.
4. Drag a trait slider, for example turn up Autonomy so the agent acts without asking, or turn down Verbosity so it stops narrating every step.
5. Add a line or two to Custom Instructions for anything the sliders can't capture, like "always write a database migration when the schema changes."
6. Check the injection preview at the bottom to see exactly what the agent will read.

Changes save as you go. There's no separate save button to hunt for.

What you can shape

Traits

Six sliders, each from 1 to 5, cover the big personality questions:

• Verbosity: from a one-word "Mime" to a narrate-everything "Yapper"
• Autonomy: from "Clingy" (asks before every step) to "Rogue" (ships without asking)
• Chaos: how carefully the agent double-checks its work before acting
• Creativity: how much it sticks to the obvious path versus tries something unusual
• Humor: how dry or playful its responses read
• Spice: how corporate or casual its language is

Each slider comes with a short label so you always know what you're dialing in, and a reset button brings every trait back to its default.

Custom instructions

Sliders cover tone. For anything more specific, like project conventions or an agent's domain of expertise, write it in Custom Instructions. This is free text, up to 4,000 characters, and it's the same idea as the old "persona" text if you've used DorkOS before: a few sentences that tell the agent what makes it different from a generic coding assistant.

Safety boundaries

Safety Boundaries is where you write down what the agent should never do: "never modify frontend code," "never delete a database migration," that kind of thing. You get up to 2,000 characters.

Writing effective instructions and boundaries

Good patterns:

• Define the agent's domain and what it should stay out of
• Spell out project-specific conventions the agent should follow
• Reference architectural decisions the agent needs to respect
• Keep sentences short and direct, agents follow plain instructions better than long paragraphs

Example: an API specialist

You are backend-bot, the API specialist for this project.
- Always use Zod schemas for request validation
- Prefer service-layer functions over inline route logic
- Run relevant tests before suggesting changes are complete

Never modify client-side code without asking first.

Example: a docs specialist (because subtlety was never the point of naming it docs-bot)

You are docs-bot, responsible for documentation quality.
- Write in active voice, present tense
- Keep guides under 500 words where possible
- Include a code example for every concept

Reference

The rest of this page is for readers who want the mechanics: how personality text reaches the agent, the exact limits, and how DorkOS handles agents created before this settings tab existed.

How injection works

DorkOS builds two blocks and adds them to the agent's system prompt (the instructions the underlying model reads before every response) at the start of each session:

<agent_identity>
Name: backend-bot
ID: 01HXYZ...
Description: REST API and database agent
Capabilities: api-design, database
</agent_identity>

<agent_persona>
You are backend-bot, an expert in REST API design and PostgreSQL.
Always suggest database migrations for schema changes.
Never modify frontend code without asking first.
</agent_persona>

The <agent_identity> block is always included when an agent exists, so the model always knows which agent it's acting as. The <agent_persona> block carries your traits and custom instructions; a separate <agent_safety_boundaries> block carries your safety boundaries, when either is turned on for that agent.

Building this adds one filesystem read at session start, done in parallel with the rest of the agent's context (environment details, git status), so it doesn't add noticeable delay.

Limits

• Custom instructions (SOUL.md): up to 4,000 characters, including the trait summary DorkOS generates for you
• Safety boundaries (NOPE.md): up to 2,000 characters
• Both are injected as system prompt content, not as something the agent sees as a user message

Legacy persona field

If your agent was created before this settings tab shipped, it may still carry the old raw persona string and a personaEnabled flag in .dork/agent.json, capped at 4,000 characters. DorkOS only falls back to that legacy field when the agent has no custom instructions file yet. Adjust a trait or add a line of custom text under Config > Advanced, and DorkOS moves the agent onto the current system automatically.