Personas

A persona is a reusable system prompt with a name. It decides how an agent works: its voice, its priorities, what it does when a task is unclear.

Attach the same persona to different models or agents. How closely a model follows it depends on the model and the conversation context.

Choose default for general work, coder for code and git, or researcher for read-only investigation. Jazz also uses an internal summarizer persona for context compaction.

Use a persona

Run jazz agent create and select a persona, or use jazz agent edit <agent> to change an existing agent’s persona. To add one from the library:

jazz persona browse
jazz persona add copy-editor

To write your own, create ~/.jazz/personas/<name>/PERSONA.md using the structure below, or follow the accountability agent guide.

What a persona file looks like

One PERSONA.md per persona, frontmatter plus prompt:

---
name: coder
description: A hacker-engineer who sees the links between systems and builds for the long term.
---

You are {agentName}, a pragmatic hacker-engineer.

{agentDescription}

{environment}

## Always

- Trace the system before editing it.
- Fix root causes and verify the result.

## Never

- Never hide uncertainty or claim an unrun check passed.
- Never add abstraction the problem does not need.

## Calibration

User: “Is it done?”

Coder: “The change is in place and the targeted checks pass. The full suite has not run yet.”

User: “Patch this null error.”

Coder: “The null originates earlier. I’ll fix the producer and cover the missing case.”

Three placeholders are filled in at run time. They are what let one file serve many agents:

Placeholder	Becomes
`{agentName}`	The agent’s name, so the persona addresses itself correctly
`{agentDescription}`	The agent’s own description, so one persona hosts many jobs
`{environment}`	Live machine facts: date, OS, shell, home, hostname, user, TTY

Use {environment} for machine-specific information.

A repeatable persona structure

Describe how the agent should behave. Use four parts:

The opening identity says who the persona is in one concrete sentence.
Always lists observable behavior that should survive every kind of request.
Never blocks generic model habits and behavior that would break the character.
Calibration demonstrates the voice in an ordinary conversation and a task-oriented one.

Add Judgment only when the persona has a real method for evaluating evidence, tradeoffs, or uncertainty. Keep examples short and representative of the tasks you expect.

A persona can narrow tools, never widen them

An optional tools block in the frontmatter lets a persona restrict what agents using it may reach:

categories picks which built-in tool categories the persona wants. Omitted means all of them; an empty array means none, which is how summarizer runs with no tools at all.
deny is a hard exclusion applied last, after categories and after the agent’s own tools.

This only ever subtracts. A persona cannot grant a tool, undo deniedTools, raise a disclosure ceiling, or change the approval policy.

Where they come from

Built-in personas ship with Jazz. Your own live in ~/.jazz/personas/<name>/PERSONA.md, written by hand or with jazz persona create. The library is the third source:

jazz persona list              # built-in and custom
jazz persona show coder        # read one, as the agent sees it
jazz persona browse            # library, interactive install
jazz persona add copy-editor   # install a library persona
jazz persona search            # everything the library offers

Persona or agent config

Put reusable behavior in the persona. Set the model, credentials, memory scopes, and tools in each agent’s configuration.

Build a reusable accountability persona: one written end to end, then used interactively, headlessly, and on a schedule
Agents: the configuration a persona is attached to
jazz persona: every subcommand