Welcome back. Let’s demystify a few terms you’ll see as you build: models, tokens, and cost.

A model is the engine behind your agent — the AI system that actually reads your instructions and generates the output. Different models are trained differently, so they behave differently: some are faster and cheaper but less precise, others are slower and more expensive but handle complex reasoning better.

That means model choice matters. A simple task, like drafting a quick reply, doesn’t need the most powerful model available. A complex task, like analyzing a long document for subtle issues, benefits from a stronger one.

Now, tokens. A token is roughly a chunk of a word — AI models read and write in these chunks, not whole sentences at once. Every prompt you send and every response you get is measured in tokens.

Here’s why that matters: cost is based on tokens used. More context, longer instructions, and longer outputs all mean more tokens, which means more cost. This is also why unnecessary information in a prompt isn’t just messy — it actually costs you, every single time your agent runs.

You don’t need to calculate this by hand. But understanding it helps you make good decisions: pick the right model for the task, and keep your instructions and knowledge base focused, not bloated.

Next, let’s put this into practice by learning how to give AI clear instructions through chat.

Welcome back. Let’s talk about giving AI clear instructions — through chat, the way you’ll actually use it every day.

Start with clear instructions. “Write me a report” is vague. “Write a one-page weekly report summarizing support tickets, grouped by issue type, in a table” is something AI can actually act on, right in the chat window.

Next, context. In that same message, give AI what it needs to do the job well — who it’s for, what it’s based on, any relevant background. But more isn’t always better. Pasting in every document you have, just in case, doesn’t improve the result — it just adds tokens, and cost, without adding value. Share what’s relevant, and leave out what isn’t.

Then, tell it exactly what output you want. What format — a list, a table, a short paragraph? What tone? How long? The clearer you are in chat, the less back-and-forth you’ll need.

If the first reply isn’t right, that’s normal — chat is built for this. Reply back with exactly what to fix, instead of retyping your whole request from scratch. That single correction is usually faster than starting over.

This works the same whether you’re chatting with AI directly, or writing instructions for an agent. Specific and lean beats long and vague, every time.

Next, we’ll look at what your agent needs beyond a good conversation: tools and knowledge.

Welcome back. Let’s break down the three things that actually shape what your agent produces: tools, knowledge, and skills.

Tools are the apps you already use every day — email, HubSpot, Notion, Slack, a spreadsheet. Connecting a tool lets your agent actually act inside it: send an email, create a HubSpot record, update a Notion page. No tool connected, no real action — it can only describe what it would do.

Knowledge is different. It’s the reference material your agent checks before answering — a policy, a price list, an FAQ, past reports. Knowledge shapes accuracy: give it the wrong or missing knowledge, and it’ll guess, or sound confident while being wrong.

Skills are different again. A skill is the process you teach your agent — the steps you follow every time, the way you want a task done. Skills shape consistency: without one, your agent might do a task differently each time, or skip a step you always expect.

Put together: tools control what your agent can do, knowledge controls what it knows, and skills control how it does the work.

Now let’s connect them. In your agent’s settings, connect the tools you use daily — the ones you listed before this lesson. Upload your knowledge document. Then add a skill: write your process out step by step, the way you’d train a new hire, and save it.

Do all three, and your agent stops guessing, and starts working the way you actually do.

Next, we’ll cover how to check that everything it produces is actually right.

Welcome back. Let’s talk about something every AI user runs into eventually: hallucinations.

A hallucination is when AI states something false with complete confidence — a made-up fact, a citation that doesn’t exist, a number that’s just wrong. It doesn’t sound uncertain when this happens. That’s what makes it dangerous: confidence is not the same as correctness.

So, how do you check? Start simple. Verify anything factual against a source you trust — a document, a number, a policy. If AI cites something, don’t assume the citation is accurate; check it. If there’s a calculation involved, do a quick sanity check yourself.

Now, when does this need a full human review, and when doesn’t it? Low-stakes, easily-checked output — like a first-draft summary you’ll read anyway — needs light verification. High-stakes output — anything that goes external, affects money, or is hard to undo — deserves a closer human review every time, at least until you’ve built real trust.

That trust should be earned gradually. When you’re testing a new agent, verify closely. As it proves accurate on a specific task, again and again, you can ease up — but only for that task, based on real evidence, not assumption.

This habit connects everything else in this course. If output looks wrong, trace it back — is the knowledge base missing something, or was the prompt unclear? That’s how you actually fix it, not just notice it.

Next, let’s build your first agent from a template.

Welcome back. Time to put everything together and build your first agent.

Leapforce gives you over 100 ready-made templates, covering common jobs across the business. Every template comes with a dashboard, automation, and report already built in — you’re starting from a working agent, not a blank page.

Start by typing what you need in the search bar, or browsing the template library, and picking one close to your real task. It doesn’t need to be a perfect match — just the closest starting point.

Once you’ve picked one, look at what’s already built in. Understand the dashboard, the automation, and the report before you change anything — you’re customizing a working setup, not starting from zero.

Then connect what it needs. Remember knowledge and tools from earlier in this course? This is where you use them — attach the document or policy your agent should reference, and connect the tool it needs to actually take action, like a live data source or system.

Now customize it for your workflow. Do this by chatting with the agent directly — the same way you’d ask an assistant for something, then approve it before it’s done. Tell it what’s different about your task.

Give your agent a name your team will recognize, and you’ve got a real, working first draft.

Next, we’ll test it properly, fix what’s off, and get it ready to run for real.

Welcome to the last lesson in this course — where your agent goes from a first draft to something you can actually rely on.

Start by testing it against real scenarios: the examples you prepared before this lesson. Include a typical case, and at least one edge case — something messy or unusual. Run each one and actually read the output, the way you’d review a new hire’s first week of work.

Here’s the good news: everything from here happens right in chat. If you spot a failure, a mistake, or a gap — the agent missed a step, used the wrong information, or got the tone wrong — just tell it, in plain language, the same way you’d correct a teammate. “You missed step three” or “this should use the updated pricing” is enough.

Your agent updates its instructions based on that conversation. Then test the same scenario again, right there in chat, to confirm the fix actually worked — not just that something changed.

No code, no separate settings screen. Identifying the gap, fixing it, and validating the fix all happen in the same conversation.

Once your test cases hold up, it’s time to put your agent to work. Publish it, then share it with your team — start narrow, with the people closest to the task, rather than opening it to everyone right away. Let them know what it does, and what still needs a human.

That’s the full arc: you understood what AI and agents can do, learned to give clear instructions through chat, connected tools, knowledge, and skills, learned to verify results, and now you’ve tested, improved, and published a working agent — all through conversation.

You’re ready to keep building. Congratulations.