Skip to content
Get packs

Guide · Agent skill

How to use the Soul Audit agent skill (soul-audit)

Updated · written with AI from the project's own documentation and checked against our catalogue data

Soul Audit is an agent skill that lets an AI agent update the files describing who it is, who you are, what it may share, and how often it checks in. It is built for people who already run the gbrain agent system and have finished first-time bootstrap. After this guide you will know what the skill does, how to install it, how to trigger it with plain requests, and what to check before and after it changes anything.

Soul Audit Identity Buildergarrytan/gbrain · MIT

Quick start

  1. 1Check requirements. Confirm the gbrain command-line tool with bootstrap interview and render subcommands is installed, and that your agent can run shell commands.
  2. 2Install the skill. Download the ZIP from this page and unzip it into ~/.claude/skills/ or a project's .claude/skills/ folder, or upload it under Settings → Capabilities → Skills.
  3. 3Start a request. Ask in plain words, for example "change my agent's personality" or "soul audit".
  4. 4Review current answers. The agent runs gbrain bootstrap interview --show so you can see what the answer bank holds now.
  5. 5Record and read back. Update the answers you want changed, then have the agent read the full set back to you and check the status hash.
  6. 6Confirm and re-render. After you agree, the agent confirms the hash and re-renders the affected file with gbrain bootstrap render --only <FILE> --force.

What an agent skill is and what soul-audit does

An agent skill is a folder containing a SKILL.md file that tells an AI agent when and how to carry out a specific task. This one is named soul-audit and is version 2.0.0 in its source file. Its job is to re-run or deepen the agent's identity interview after the first setup.

The skill does not write identity text itself. It asks you questions in short phases, records your answers word for word into a shared answer store (state/interview.json), reads them back to you for confirmation, and then re-renders the affected files with the gbrain bootstrap render command. The files it can feed are SOUL.md, USER.md, ACCESS_POLICY.md, HEARTBEAT.md and AGENTS.md. The source states that the agent never invents or defaults identity answers on your behalf.

  • Phase 1: identity (agent name, purpose, relationship) feeds SOUL.md
  • Phase 2: vibe calibration (voice, what made you wince) feeds SOUL.md
  • Phase 3: mission (top goals) feeds SOUL.md and AGENTS.md
  • Phase 4: user profile feeds USER.md
  • Phase 5: boundaries (who can see what) feeds ACCESS_POLICY.md
  • Phase 6: operational cadence (check-ins, quiet hours) feeds HEARTBEAT.md

Install the soul-audit skill

Download the ZIP from this directory page and unzip it. The skill is a folder that contains SKILL.md. Put that folder in one of two places: your user skills folder at ~/.claude/skills/, or a project-specific folder at .claude/skills/ inside your project.

If you use the Claude apps instead, you can upload the skill there under Settings → Capabilities → Skills.

The skill needs the gbrain command-line tool with its bootstrap interview and render subcommands, and a shell tool the agent can run commands with. If you do not yet have a bootstrap workspace, the skill can start one with gbrain bootstrap interview --init.

  • Option A: unzip into ~/.claude/skills/ for all projects
  • Option B: unzip into .claude/skills/ inside one project
  • Option C: upload through Settings → Capabilities → Skills in the Claude apps

Trigger the skill with natural requests

You do not need to remember a command name. The skill's source lists these phrases as triggers, and you can use wording close to them. Say what you want changed and the agent will start the interview.

Start with a narrow request if you only want one thing updated. The phases are independent, so you can skip straight to one of them.

  • "soul audit"
  • "customize agent"
  • "who am I"
  • "set up identity"
  • "change my agent's personality"

Example: update one answer, such as quiet hours

Suppose you only want to change when the agent stays quiet. You would ask something like: change my quiet hours. The agent should read the current answers with --show, record your new QUIET_HOURS value verbatim, read the full set back to you, and confirm only after you agree.

The skill then re-renders the file that depends on that answer, using the same render path as first-time setup. The rest of the interview is not repeated.

gbrain bootstrap interview --show gbrain bootstrap interview --set QUIET_HOURS "your own words here" gbrain bootstrap interview --status gbrain bootstrap interview --confirm <hash> gbrain bootstrap render --only HEARTBEAT.md --force

Read the change log and backups

The skill reports which keys were updated, which files were re-rendered, and where backups went. Rendering with --force first copies the old file to .gbrain-bootstrap-backups/ under a timestamped folder, so nothing is overwritten without a recoverable copy. Keep this in mind if you edit rendered files by hand: the next render replaces them, so change the answer instead and re-render.

In an OpenClaw workspace, the HEARTBEAT.md target writes SCHEDULES.md, because OpenClaw 2026.10.6 and later removes a workspace HEARTBEAT.md. Scheduled jobs ship disabled and should be enabled one at a time after a manual dry run.

  • Backups live in .gbrain-bootstrap-backups/<timestamp>/
  • Hand-edits to rendered files are overwritten on the next render
  • Jobs ship disabled; enable one at a time

Safety notes: scripts and network

The listing states that this skill runs scripts and does not use the network. The skill is marked as mutating, meaning it changes files in your workspace. It uses the shell tool and the gbrain commands.

Boundaries matter. The access phase controls who can see what, so a mistake there could expose private information. Review ACCESS_POLICY.md after any change, especially in a shared or multi-person workspace. The skill is not designed for setting access rules across several people at once.

  • Runs scripts: yes. Uses network: no (per the listing).
  • Mutates files in the workspace; rendered files are backed up first.
  • Review access rules carefully before sharing a workspace.

Tips for a good audit

Answer in your own words. The skill records what you say verbatim, so short honest answers beat polished ones. Skip optional questions with an explicit decline rather than leaving them vague.

Run one phase at a time. The skill's own rule is that an overwhelming all-at-once interview is worse than a short focused one. Confirm the read-back before any render, and only confirm answers you have actually seen.

  • Speak your answers; the skill does not prefill them
  • Do one phase per session when possible
  • Read the read-back carefully before confirming

What people use it for

  • Change the agent's tone from casual to direct after a few weeks of use
  • Update your timezone or current projects in USER.md without redoing the whole interview
  • Adjust quiet hours in HEARTBEAT.md so check-ins stop at night
  • Tighten who can see what in ACCESS_POLICY.md before adding a collaborator
  • Revise top goals so SOUL.md and AGENTS.md reflect current priorities

Frequently asked questions

Is the soul-audit skill free and what is its licence?

The listing states the licence is MIT. The source material does not mention pricing, so check the listing for any cost details before you install.

What do I need before I can use soul-audit?

You need the gbrain command-line tool with the bootstrap interview and render subcommands, and a shell tool the agent can run commands with. You also need a bootstrap workspace; if you do not have one, the skill can start one with gbrain bootstrap interview --init.

Where do I install the soul-audit skill?

Unzip the downloaded ZIP into ~/.claude/skills/ for all projects, or into .claude/skills/ inside one project. In the Claude apps you can upload it under Settings → Capabilities → Skills.

Will the agent write my personality or profile automatically?

No. The skill is designed so the agent never invents identity content. It asks you questions, records your answers verbatim, and only renders files after you confirm the read-back.

What happens to my old files when they are re-rendered?

Rendering with --force backs up the old file to .gbrain-bootstrap-backups/ under a timestamped folder first, so a recoverable copy exists.

Does soul-audit replace the first-time bootstrap?

No. It is the re-run and deepen surface. First-boot bootstrap and this skill write to the same answer bank, so an update made here flows into the rendered identity files.

New agent skills, repos and AI tools — every week

The best new skills and open-source projects, the fastest-growing repos, and one free prompt. Five minutes to read. Unsubscribe anytime.

Our verdict

Soul Audit suits people already running gbrain who want to change their agent's tone, goals, access rules or check-in schedule with a confirmed record of each change. It is not a good choice if you have not installed the gbrain bootstrap tooling, want the agent to fill in personal details for you, or are setting access rules across a shared workspace without careful review.

Download the skill