How to set up Minds: create your first personal AI agent

Define the job and limits first

Before choosing a template or setting, name the situation you want help with and the result you want back. Give one Mind one role, then mark the decisions and actions that must stay with you. A useful first instruction is:

Help me [complete a defined job] using [approved information or tools], return [specific output], and ask me before [restricted action].

Prepare five things:

1. Goal: the outcome the Mind should produce.
2. Context: the information it needs.
3. Format: how it should return the result.
4. Boundaries: the sources, tools and actions allowed.
5. Review rule: what you must check or approve.

Do not include confidential or personal information unless it is necessary, permitted and appropriate. Review the live privacy policy and connected-service terms before granting access.

Choose the right setup route

Choose the smallest setup that can handle the first job.

Choose and set up a Mind template

Open hellominds.ai and select Launch a Mind. Choose a template by reading its role, intended use, tools and required connections rather than choosing by name alone. In the interface observed on 11 September 2026, the next choice was Quick Start or Customize. Quick Start used recommended settings; Customize provided more control over needs and preferences. The catalogue and labels can change.

Before completing setup, check whether sign-in is required, whether the Mind or configuration has a price, how Cognition is used, which apps it wants to access and whether a recurring payment is selected. Give only the access required for the defined job.

In the observed Quick Start flow, you name the Mind and select Launch Now. The completion screen says the Mind will send a greeting email and can be connected through Telegram. Confirm the name and role, instructions, permitted capabilities, expected channel and available controls before starting important work.

Create your own Mind

Select Create Your Mind when no current template matches the job. Give the Mind a role you can picture in your week, such as “weekly research analyst” or “project follow-up coordinator”, and avoid titles broad enough to mean anything. State the result rather than only the activity: for example, ask for a Friday briefing containing sources, publication dates, what changed, why it matters and any uncertainty.

Add working instructions for permitted sources, required language and tone, output structure, evidence citations, conflict handling and when to stop and ask. Then set boundaries such as:

• do not publish, send, purchase or change access without approval;
• do not use personal data outside the defined task;
• do not present uncertain information as fact;
• do not provide regulated advice.

Review the generated configuration for assumptions before creating the Mind. Check the current price, Cognition balance or purchase option, account requirement and terms, and keep a copy of the intended job definition.

Review access and Circle permissions

The current completion screen warns that people added to a Circle may interact with the Mind and assign work without advance approval. Their activity may use Cognition. Circle members may also see information the Mind knows about you and request Cognition transfers on your behalf. Add only people you know and trust, review the Mind's access, and confirm the current Circle controls before sharing it.

Test the first task

Use a representative task whose answer you already know well enough to evaluate. Specify the objective, inputs, method, format, boundary and review point. After the response, check whether the Mind followed the source boundary, invented details, used the requested format, recognised uncertainty and stopped where approval was required. Correct the instruction and run a second low-risk test before expanding the role.

Add tools and connections gradually: start without external access, then try narrow read-only access, draft-only access and finally a reversible action with explicit rules. Remove access the Mind no longer needs.

Improve and troubleshoot

When a result misses the mark, replace “make it better” with a measurable rule, such as “keep the summary under 150 words, separate confirmed facts from interpretation, cite every factual claim, and write ‘source needed’ when evidence is missing”. Keep a short record of the job definition, approved connections, approval points, good output and known failure patterns.

If the role is too broad, split it into one job. If output varies, add a fixed template and example. If the Mind takes an unexpected step, reduce permissions and strengthen approval boundaries. If a route or connection looks different or is unavailable, check the current release, account, region and service terms rather than assuming equivalence.

Create your first Mind

Start with one clearly defined job. Test the result, then widen access only if the Mind behaves as expected.

Explore Minds or browse ready-made Minds.