For every agent

Persona and rules

A persona is the name you give an agent, the tone it uses, and the way it talks to you. Rules are standing instructions it follows on every job, even when you forget to repeat them. Memory is separate, because it holds facts the agent learns about you, such as how you like a summary written. A skill is how to do one kind of job, and Skills and MCP explains skills and the extra tools an agent can gain.

Where each setting goes

Eight of these agents keep a persona and rules in one of four places. Atomic Agent and OpenManus have no saved persona file, and the note under the table says what each one uses instead. This table names the kind only, and the exact file, folder, or screen stays on that agent's own page.

Kind of setting What you write there Agents
A persona or soul file The name, the tone, and how it talks to you OpenClaw, Hermes Agent, QwenPaw, NanoClaw, Everfern
A rules file, such as AGENTS.md Standing instructions it should follow every time A separate file. See the note under this table.
A system prompt box The same kind of words, typed into a box in the app None of the ten in the versions checked here
A profile or prompt folder A saved profile, or a folder of prompt files Agent Zero, Goose, Eigent

Atomic Agent builds the voice into the program, saves preferences as pinned profile facts, and loads AGENTS.md from a trusted folder on cloud-model chats only. OpenManus has no saved persona or rules file, so put the tone and limits in each prompt and leave app/prompt/ alone. NanoClaw keeps the voice and standing rules together in instructions.prepend.md. OpenClaw and QwenPaw keep tone in SOUL.md, the name in IDENTITY.md or PROFILE.md, and standing rules in AGENTS.md. Hermes pairs SOUL.md with project rules in AGENTS.md, AGENTS.override.md, or .hermes.md, and Everfern pairs SOUL.md with agents.md. Agent Zero puts both in an agent profile's Instructions box, Goose uses system.md and .goosehints, and Eigent stores Space role Instructions while your own saved tone and rules go in a custom SKILL.md.

How to write them

Keep the persona and the rules short and specific, and phrase each rule as something the agent should do. Put the most important rule first, because a long list is easier for the model to drop. A weaker model follows a shorter list more reliably, and Prompting strong and weak models shows how to match the words to the model you have.

A starter persona and rules

Paste these into the place your agent's own page names, and change the name if you want a different one. Leave the three rules as they are until you have seen the test question work.

Name: Riley
Tone: calm and plain

Talk to me in short sentences and plain words.
Skip greetings and filler.
When a job has two sensible choices, ask which one I want.
1. Ask me before you delete, move, or overwrite any file.
2. When you are unsure, say so in one sentence and wait.
3. Read and write files only inside the folder I name for this job.

Rules are not a lock

The words are only a request, so a model can ignore a rule and do something else. A web page or a file can hide new instructions that try to override your rules, which is called prompt injection. Approvals, a named folder, and a container are what limit the agent, and those stay in place when the words fail.

Check that they took effect

  1. Save the persona and the rules, start a new chat, and ask the question below.

What is your name, and what will you do before you delete a file?

What you should see it answers with the name Riley and says it will ask you before it deletes a file.

A short persona and a few rules.

Sources

Checked 14 November 2026.

Next

The next page is After it does something wrong.

Checked 14 November 2026 · prices, limits, and versions change