Tutorial: build a browser-automation Digital Human

Tutorial: turn Halo's OA approval sample into a Digital Human that works in a web system you sign in to, with bundled browser scripts, memory and safe defaults.

In this tutorial you adapt a sample into a Digital Human that works inside a web system you sign in to. The sample is an approval assistant for an office system: every 30 minutes it lists your pending approvals, handles the ones that match your rules, and reports the rest. Along the way you meet every part of a production Digital Human: the spec, bundled browser scripts, memory, a login, and a safe default.

Before you start

• Halo with an AI model set up.
• A web system you sign in to in a browser, such as an approval, ticket or CRM system, whose pages you can inspect with the browser's developer tools.
• Some JavaScript, if you want to adapt the scripts by hand. Halo can also adapt them for you.

1. Download the sample

Download oa-approval-agent.zip and unzip it into your workspace folder:

oa-approval-agent/
├── spec.yaml                 the Digital Human
├── README.md
└── skills/
    ├── oa-todo-list/         script: list pending approvals
    ├── oa-approve/           script: approve or reject one
    └── oa-approved-list/     script: list approved items, to double-check

The sample does not run as it is

Its addresses, API paths and field names are placeholders such as your-oa.example.com. You replace them with your system's in step 4.

2. Read the spec

Open spec.yaml. Five parts matter:

• requires.skills lists the three scripts as bundled skills, so they are installed with the Digital Human into its workspace's .claude/skills/, the paths the prompt's browser_run calls use.
• system_prompt is a procedure: a list of forbidden actions, then numbered steps that each name the exact call, such as browser_run({ file: ".claude/skills/oa-todo-list/index.js", params: { pageSize: 30 } }), and what to do when it fails. The last step reports with report_to_user.
• config_schema is the settings form. auto_approve is a boolean that defaults to false, so the Digital Human only reports until you turn it on. max_approvals_per_run caps how many it may handle in one run.
• memory_schema keeps processed_todos, so an item is never handled twice, plus the time of the last run and running totals.
• browser_login names the system's address, so Halo asks you to sign in there.

Each field is described in the spec reference.

3. Read a browser script

Open skills/oa-todo-list/index.js. It is one async (params) => { … } function that runs inside the signed-in page:

async (params) => {
  const { pageSize = 30 } = params || {}
  try {
    const resp = await fetch(`https://your-oa.example.com/api/todo/list?pageSize=${pageSize}`, {
      credentials: 'include',
      headers: { Accept: 'application/json' },
    })
    if (!resp.ok) {
      return { success: false, status: resp.status, error: resp.status === 401 ? 'not logged in' : `HTTP ${resp.status}` }
    }
    const data = await resp.json()
    return { success: true, todos: data.list ?? [] }
  } catch (e) {
    return { success: false, error: e.message }
  }
}

The script calls the system's own API with the session you signed in with, and it never throws: every outcome comes back as { success, … }, so the prompt can tell "not logged in" from other failures. This is shortened from the file; the full rules are under Browser scripts.

4. Adapt it to your system

Find the requests your system makes. Sign in to it in a browser, open the developer tools on the Network tab, filter for Fetch/XHR, and approve or reject one item. Note each request's URL, method, body and the fields of its response.

Then adapt the package in one of two ways:

• Let Halo do it. In a conversation in this workspace, say: "Adapt the Digital Human in the oa-approval-agent folder to our approval system at <YOUR_SYSTEM_URL>. The list of pending items comes from <LIST_REQUEST> and approving one sends <APPROVE_REQUEST>. Update the three scripts, browser_login and the address in the first step of system_prompt, and keep the rest of the spec."
• By hand. In each index.js, replace the base address, the API paths and the field names. In spec.yaml, set browser_login[].url and the address in the first step of system_prompt.

If an action needs a token from the page, such as a CSRF token in a form field, have the script read it from the page before it calls the API.

5. Import it and sign in

1. In Halo, open Digital Humans, click Create Digital Human, choose Import, and drop the oa-approval-agent folder (or a .zip of it). Import installs the bundled scripts along with the Digital Human.
2. Open the new Digital Human, go to Capabilities and settings, and fill in the keyword settings. Leave auto_approve (labelled 自动审批 in the sample) off.
3. Above the Digital Human's Work activity, click Open to log in for your system and sign in.

6. Test it

1. Click Run now on the Digital Human's page.
2. Open Work activity and read the run's report. With auto_approve off it only suggests: it lists what it would approve and reject, and changes nothing in your system.
3. If the report says it is not logged in, sign in again with Open to log in. If a script fails, its error is in the report; debug it as described under Develop and debug a script.
4. When a few reports match what you would have decided, narrow the keywords to a small, safe set and turn auto_approve on.

Done

The Digital Human now checks your system every 30 minutes on its own, handles only items that match your rules, never more than max_approvals_per_run at once, and reports the rest to you.

7. Share it

In the desktop app, on the Digital Human's page, click ⋮ › Share › Export to get a .dhpkg package, or submit it to a Store, where it is listed after review. See Publish to the Store. Remove company addresses and names before you share it outside your company.

Troubleshooting

What you see	What to do
The report says the OA login has expired (HTTP 401)	Sign in again with Open to log in. Sessions expire on the system's side
Path not allowed	The script is outside the workspace and its .claude/skills/ folder. Import the package with Create Digital Human › Import so the skills land in the workspace
The Digital Human doesn't run on its own	Check subscriptions (the sample runs on every: "30m") and that it isn't paused. Scheduled runs are also skipped while one of its questions is unanswered, such as its request to sign in again: answer it in Work activity
Items are handled twice	The last step must write the handled ids into memory before the report

Next steps

• Production-grade Digital Humans: prompts, memory and scripts that hold up
• Digital Human spec reference: every field
• DHP protocol: how a package is laid out