AISquare
rlx

Personas: pick, edit and import

Run any rlx agent as someone, such as a skeptic, a mentor or a minimalist. Pick a persona when you spawn, edit one in your own editor, import your own skills or plain text, and attach one to an agent that is already running.

A persona is how an agent works, written down once and given to it when it starts: a skeptic that reproduces before it believes, a minimalist that makes the smallest change that meets the contract. The role still decides what the agent does. A coder running as skeptic is still a coder, with a coder's work cycle and its task's contract.

A persona is a Claude Code skill: a directory named for the persona, holding a SKILL.md, so there is no rlx format to learn. It is --persona skeptic to the fleet; to use it as /skeptic in Claude Code too, copy it into Claude Code's skills with rlx persona export skeptic --skill (with --user or --project).

Where personas come from

rlx looks in three layers. When two hold the same name, the higher one wins.

LayerWhereWho sees it
project<project>/.aisquare/personas/<name>/This project, and anyone who checks the repository out
user~/.aisquare/personas/<name>/You, in every project on this machine
bundledinside rlx, read-onlyEveryone

The app shows your user personas and the bundled ones, the same in every project, and writes only the user layer. A project's own personas are read and written from a shell, with --project.

Four personas ship bundled:

PersonaHow it works
carefulPrefers reversible steps, states assumptions before acting, and asks one precise question when blocked rather than guessing.
mentorExplains the why before the what, surfaces the alternatives it weighed, and leaves the reader able to do it next time.
minimalistThe smallest change that meets the contract. No speculative abstraction; says what it left out and why.
skepticEvidence first. Treats a green check as a claim, reproduces before believing, and names what it did not verify.

An agent with a persona is briefed with the persona's body, then one sentence rlx adds: the persona shapes how the agent works and communicates, and never overrides its role's cycle, the lane rule, a task's contract or evidence. The body counts against every turn's context, so a body over 4,000 characters gets a warning and one over 12,000 is refused.

Run an agent as a persona

Open the Spawn dialog

In the sidebar, press + Spawn agent, and pick the Project, the Role and, if you want, the Account.

Pick a persona under Advanced

Open Advanced. The Persona select lists your personas and the bundled ones, with the chosen one's description under it. It opens on the role's own default, if the role has one; (none) runs the agent with no persona, even over a role default. Import… beside it imports a persona and selects it.

The Spawn dialog's Advanced section with Persona set to test-writer, its description under the field.

Spawn

Press Spawn. The agent's info line, at the top of its view, and its row in the sidebar show the persona it runs as. A restart or an account switch keeps it.

Give a role a default persona

To run every agent of a role as a persona, open ⚙ Settings ▸ App settings, then Agent settings ▸ Advanced, and pick it in the role's persona select. It lists your user personas and the bundled ones, which every project can resolve; a project's own persona is given per spawn from a shell, with --persona. Press Save.

From a shell, the same default is:

rlx config set fleet.roles.reviewer.persona skeptic

A spawn's own choice wins over the default. A default that a project cannot resolve refuses the spawn, naming the setting.

From a shell

rlx fleet spawn coder --persona minimalist --prompt "Start on #12"
rlx launch coder --persona skeptic       # a session you run by hand

Settings ▸ Personas

Your personas live in ⚙ Settings ▸ Personas, the same page for every project. Open ⚙ Settings at the bottom of the sidebar, then Personas.

⚙ Settings ▸ Personas: your personas and the bundled ones in one searchable list, with the exact briefing a session gets for the highlighted one.
  • Search by name, description or tag, and narrow the table with the user and bundled layer filters.
  • The preview on the right shows exactly what an agent would be briefed with, for the selected row.
  • Actions under the table: Edit, Export…, Remove and Validate. With the keyboard on the page: e edit, x export, Del remove, v validate, i import, and Enter opens the selected persona's SKILL.md.

A bundled persona cannot be edited or removed in place. It opens read-only, with Save as… to copy it into your user layer, where you can change it.

Edit a persona in your own editor

Select it and press Edit

Select one of your personas and press Edit (or e).

Edit in the pane

Your editor opens in a pane in place: $VISUAL, else $EDITOR, else vi. It works on a copy of the SKILL.md, never on the file itself.

Edit opened a persona's SKILL.md in vi, in a pane inside rlx; save and quit, and it is checked and saved.

Save and quit

When you save and quit, the copy is checked and saved, and the table reads the layers again. A copy you did not change writes nothing. Text that does not parse is not saved: the dialog says why, and Reopen editor opens the same copy again, so nothing you typed is lost.

On Windows, on a machine without tmux, or with no editor to run, Edit opens a text box inside rlx instead, with Save and Cancel.

Import a persona

+ Import… (or i) turns something you already have into a persona.

Name the source

In Source, give a skill directory, a SKILL.md, a Markdown file, or the name of a Claude Code skill; Browse skills lists the skills Claude Code already has. The Layer is user, the one the app writes, and the Name is optional.

The Import a persona dialog: Source, Browse skills, Layer, Name, Use the LLM, Condense, Engine, Model and Force.

Import

Press Import. A Claude Code skill is copied in byte for byte. Anything else (plain text, another tool's persona) is converted into a skill by an LLM engine: by default headless Claude Code, else the Anthropic API when the llm extra is installed. Use the LLM, on by default, allows that conversion; switched off, nothing spends a token and a source that is not a skill is refused. Condense rewrites the body shorter.

Check the draft, then save it

A converted persona is shown to you before it is saved. Save keeps it; Discard saves nothing, though the draft stays under ~/.aisquare/personas/.drafts/<name>/, where rlx persona import ~/.aisquare/personas/.drafts/<name> can take it up again. Cancel while an engine is still working stops it within a couple of seconds.

+ New makes a persona from a template: give it a name (lowercase letters, digits and single hyphens) and a one-line description, and press Create; it goes in your user layer, and the editor opens on it next. Export… copies a persona into a coding agent's personal skills (Claude Code's or Codex's), or a directory of your choice.

Give an agent a persona

To start a new agent as a persona, pick it in the Spawn dialog under Advanced ▸ Persona (above).

To give one to an agent that is already running, use a shell:

rlx persona attach mentor --to coder-auth

An agent waiting at its prompt is briefed now; a busy one gets the persona privately with its next prompt. It stays on the agent's record, so a /clear or a restart briefs it again.

From a shell

Every action has a command. --user and --project choose the layer where one applies.

CommandWhat it does
persona listEvery persona in every layer, and any directory that did not load, with why. --tag filters.
persona show <name>Exactly what a session would be briefed with, then where it came from and its files.
persona new <name>Make one from the template and open it in $EDITOR (the user layer by default).
persona edit <name>Edit a project or user persona in $EDITOR; text that does not parse is not saved.
persona rm <name>Remove a project or user persona.
persona validate <path>Check a directory against the persona rules: errors exit 1, warnings do not.
persona import <source>…Import one or more sources (see below).
persona export <name>Print the SKILL.md, or write the directory elsewhere (--to DIR), or into Claude Code's skills (--skill).
persona attach <name> --to <label>Give a running fleet agent a persona.
rlx persona list
rlx persona show skeptic
rlx persona import ./my-skills/reviewer-voice              # a skill: copied as it is
rlx persona import notes/tone.md https://example.com/persona.md --project
rlx persona import --from-dir ~/team-skills                 # every skill and text file in it
rlx persona attach mentor --to coder-auth

persona import takes a skill directory, a SKILL.md or any text file, - for stdin, an https:// URL, or the name of a Claude Code skill (--list shows those). Give it several sources, or --from-dir DIR, and it imports them one after another into one table of results (source, name, engine, outcome): one refusal does not stop the rest, and the command exits 1 if any was refused. --from-dir lists a README, LICENSE or CHANGELOG beside your skills as skipped and spends no engine on it. --name names a single import, --llm or --no-llm forces or forbids the conversion, --engine picks auto (the default), manager, api or off, and --yes saves a draft without asking. Ctrl-C stops the import and any engine it started, writes nothing, and exits 130.

On this page