Agent Basics — Startup

Adapted from the DGF Agent Basics series — consumer-first guides for working with a personal AI agent. Pick only what you need.

> Pick only what you need. This skill is individual. This is the foundation skill — the only skill a new user truly needs, because it teaches them to build their own skill set.

Summary: The only foundation skill. How to set up your own agent, bring your own accounts, plan first, check understanding, work in micro-steps, and create your own skills and tasks by showing it once and saving by name. This one alone is enough to build your own skill set.

Use this skill on its own, as a different skill from any other skill, when a person is setting up your agent for the first time. It teaches startup, setup, and how to use your agent, and how to create a new skill or a new task. An agent that loads only this skill must be able to follow it with no knowledge of any other skill, and no assumption that skills were ever together.

Use any agent — you are not locked to one product

These skills work for just about any personal agent, not for one product only. Pick the agent that feels right for you, and that fits the work you actually do.

Every agent has limitations, and limitations hit different people differently. An agent that works well for one person, in the way that person works, may not work as well for another person, for other work. That is normal, and it is not a reason to lock someone in at the start.

Use an agent for what it is best at. If an agent cannot do a step in a skill, use the closest equivalent that agent does have, or do that step yourself, and keep the rest of the method. Do not expect one agent to do everything, and do not judge a skill by an agent trying to do work it is not good at.

If a skill names a feature your agent does not have — for example a daily feed, a suggestions feature, a browser you can take over, or a secure sign-in — use the closest feature your agent does have, or do that step by hand, and say plainly what was skipped. The method still works.

The author of these skills has a personal recommendation for an agent, because of how that agent delivers and does things, but that recommendation is not a requirement, and it is not in these skills, so no one is locked in by picking up a skill.

Get your own personal agent

Start by getting your own personal agent, from that agent's website or app store, and setting it up. Your agent is yours alone, strictly personal.

Bring your own accounts and logins. Use nothing that belongs to anyone else. No account, password, or payment method from another person carries over, and no one else's agent is yours.

Tell your agent your name, in plain words, and how you want to be called. That is the start of it knowing you.

Say plainly what your agent is not, when it helps. It is not a person, and not a company you hired. It is your personal agent, that you talk to in chat, in the way that agent runs — on a phone, on a computer, or in the cloud, depending on the agent.

Tell it what you care about

Tell your agent what you care about in plain words. Name your business, your work, your family, your hobbies, your health, your money, or any topic you want it to follow. Talking is the setup. You do not have to set up a system first, fill in a form, or learn any technical words.

Give one interest at a time, in plain words, and check that your agent heard it right. If it repeats back something different from what you meant, correct it right there.

Tell your agent what you do not want, as well. Name topics to leave alone, times not to bother you, and things you never want it to do without asking you first.

Do not lead with virtual machine or VM, and do not call it a VPN. Say it in plain benefits, in the way that agent works: you talk to it in chat, you do not have to set up a system first, it keeps your logins safe where it supports secure sign-in, and it saves and remembers your stuff.

Load your accounts safely

Load each account you want your agent to work inside, one at a time, through the secure sign-in your agent shows you. Logins go in secure storage, never in chat. Your agent never sees the values you type into secure sign-in.

Never type a password, a bank login, a card number, or a one time code into chat. If your agent ever seems to ask for one in chat, stop, and ask it to show you the secure sign-in instead.

If a login was not loaded securely, your agent works with public information only, and does not look inside that account. That is a protection, not a failure.

Check each connection after you load it. Ask your agent to show you what it can see in that account, and what it cannot see, before you ask it to do work inside it.

Ask for the plan first

Ask for the plan first, before any work runs. Say the task, then ask for the exact steps it will take, what it needs from you, and anything it cannot do. Wait for you to approve or correct the plan.

Review the plan like a checklist. Fix a wrong step, add a missing step, or remove a step you do not want, before work starts. A plan you approved is a plan you can hold your agent to.

Check that your agent understood you, before it runs. Ask it to show back what it thinks you mean and what output it will give you. Correct its repeat back right there when it does not match what you pictured. This check is called a mutual understanding check, and it prevents most mistakes.

Work in small single steps

Work in small single steps. Count the steps first. Give one step, check it, then give the next step. A ten step run is a huge run for an agent, so break it down and keep each step to exactly one thing.

After each step, check the result in a form you can actually see, such as a preview, a list, a draft, or a file. A step with nothing to check is not finished.

Do not let your agent run a long chain on an unanswered question. If it needs a fact, a choice, or a login from you, it must ask and wait, not guess.

Create a new skill or a new task

Show it once, then save it by name. This is how you create a new task or a new skill, and it is the most important setup method in this skill.

Take over the browser and do the task once yourself while your agent pauses. You drive alone and your agent does not watch your clicks or drags. There is no record button to press.

After you hand control back, tell your agent the steps in chat, in plain words, and tell it to save them by name, such as save this as Fun Things. Name it for what it does, not for how it works.

Later, say pull up Fun Things, or run Fun Things, and your agent picks up where you left off, using the saved steps. If a step no longer works, correct that step in plain words, and tell your agent to remember the correction.

Save a skill the same way, for a kind of work you repeat, such as shopping, posting, or tracking money. A task is for one job saved by name. A skill is for a kind of job, with its steps, its checks, and its limits, saved so an agent that knows nothing else can follow it.

Correct it in small plain words

Correct your agent in small plain words when it does it wrong. Do not start over. Say what to change and say to remember that correction. That preference sticks and it gets better for you specifically over time.

If your agent repeats a mistake you already corrected, say so plainly, and ask it to show you what it saved for that correction. A correction that was not saved is a correction that will be lost.

Never argue — how your agent talks

Your agent never argues. Arguing means someone was right or wrong, and that is never the way to behave with anyone, and never the way to get the best out of someone.

Your agent states facts, plainly, with source and date when a fact is decisive.

Your agent asks as if it heard correctly, asking questions based on the answers given.

Your agent offers alternative information: a different point, a different objective, or a different frame, without saying the other person was right or wrong.

Your agent asks for a better explanation: could you explain that better? Did you mean this, or did you mean this?

That is how better output is reached: alternative information, and questions that help the person explain themselves more clearly. Never by arguing, never by telling someone they were wrong, and never by telling someone someone else was right.

Let it keep learning, every day

Let your agent keep learning your interests every day. Your Feed looks like a newspaper with articles and topics related to what you told it you care about, updated daily. Your Ideas look like a lightbulb with suggestions for things you might like.

It is not mind reading. It learns from what you actually talk about. If the Feed or Ideas drift away from what you care about, tell your agent what changed, and it adjusts.

Check your Feed and Ideas once a day at first, and tell your agent which items were useful and which were not. That feedback is how it gets better for you specifically.

Use easy first, deep only when you want it

Use easy first and deep only when you want it. Never require a deep setup to use a basic task. A basic task must work with plain words, one step at a time, and no technical setup.

Ask for more depth only when you want it, such as a fuller report, a saved skill, a recurring check, or a connected account. Depth is an option, never a requirement.

What setup never does

Setup never gives your agent your money, your passwords in chat, or permission to post, publish, buy, or move money without you saying so for that exact action. A setup that is finished does not mean every later action is approved.

Setup never shares your private information with another person, another account, or another agent. If your agent is asked to share something private, it must ask you first, and name exactly what it would share and with whom.

Setup never promises unlimited free use, unlimited results, or guaranteed outcomes. every agent has limits, and this skill states its own limits plainly.

Deliverables for a finished setup

An agent that knows your name, what you care about, and what you do not want it to do without asking.

At least one account loaded securely, or a plain statement that no account is loaded yet and your agent will work with public information only.

At least one task saved by name, from a show it once run, that you can pull up again.

A Feed and Ideas check, showing that your agent is learning from what you actually talk about.

A plain list, in the same message that says setup is done, of what was set up, what was not set up yet, and what still needs you.

Diverse, not concrete: let the user prune

Do not make this skill concrete, narrow, or one size fits all. Present a diverse set of options first, across price, style, risk, format, and approach, so different people can see themselves in it.

Let the user prune. Ask what to cut, what to keep, and what to change, and then narrow to what they kept. Pruning is the user's job, not your agent's guess. Do not prune for them before they have seen the options.

Do not lock a single choice early. Offer at least two or three meaningfully different options at each decision point, with the trade for each, unless the user already locked that choice themselves.

Keep options that do not fit one person, because they may fit another person using the same skill. A skill that only works for one kind of person is too concrete to reuse.

When the user prunes, save the pruned version by name when they want to repeat it, so their version starts from what they kept, not from the full diverse set again.

Available connectors

Use your agent's connector and skill search, if it has one, to find other connectors and related skills.

If nothing here fits, look on the web for a maintained API or MCP server and set it up with your agent's custom connector setup, if it has one, or build a custom connector.