Blog › Training Your AI Employees

How to Train Marblism's AI Employees: 10 Tips From a Real Operator

By Jeff — 21-year service-business owner & Marblism partner who uses it daily
Last updated: July 2026
The one thing to understand: your Marblism agents hold information in four different places — chat (temporary), Tasks (scheduled), the Brain (permanent knowledge), and Guidelines (behavioral rules). Put the right information in the right place and everything else gets easier. Workspace says who we are, the Brain says what we know, Guidelines say how to do your job.

Most people who say "Marblism didn't work for me" made the same mistake: they typed instructions into chat, got a good result once, and expected the agent to remember forever. That's not how it works — and once you know how it does work, these agents get seriously useful. Here's what I've learned running them in my own business.

1. Chat is not memory

Anything you tell an agent in chat only lasts for that session. If your agent "forgot" something, this is almost always why. If it must survive the session, it goes in the Brain or the agent's Guidelines — not chat.

2. Write Brain memories like you're briefing a new hire on day one

The Brain accepts plain-text memories, documents, web pages, and images. The trick is specificity. Compare:

The second one gets followed. The first one gets interpreted — differently every time.

3. Many small Brain docs beat one giant one

Make separate Brain entries for pricing, brand voice, services, and rules — because you can control which agents see each item. Your social training doc should only go to Sonny; your outreach playbook only to Stan; pricing to everyone. Agents drift when you bury them in irrelevant knowledge.

4. Feed the Brain specific pages, not just your homepage

When you add a web page by URL, agents only get what's on that page. Adding your homepage alone gives them almost nothing. Add your about page, services page, pricing page, and FAQ individually.

5. Use the task formula for recurring work

Tasks are scheduled actions with context attached. The formula that works every time:

"[Agent], please create a task for yourself. [Recurrence] at [time], I want you to [specific action + details]. This is for [goal]."

Example: "Stan, please create a task for yourself. Every Monday at 9:00 AM, send a follow-up email to leads who haven't responded in 5 days. This is for keeping my pipeline active."

The goal at the end isn't fluff — the agent uses it as context when the task fires. One-time things you need today? Just ask in chat.

Use the free Marblism recurring-task prompt builder to turn those fields into one reviewable instruction. Before scheduling it, run this 12-minute Marblism timezone check so the task fires on your local clock.

6. Correct behavior in chat — it sticks

When an agent does something you don't like, tell it directly: "From now on, always…" The agent updates its own Guidelines and confirms. You can also fine-tune Guidelines manually in each employee's settings. This is the fastest way to train them — treat every miss as a coaching moment, exactly like a real employee.

7. Assign projects, not micro-tasks

Marblism agents collaborate. Give one agent a project that needs multiple skills and it pulls in teammates — ask Sonny for a social campaign and he'll grab insights from Penny's latest blog posts on his own. Try "create a blog post and turn it into a LinkedIn campaign" and watch them coordinate.

8. Set Sonny's Tone Link on day one

In Sonny's settings there's a field called Tone Link — a URL he analyzes to learn your writing style, vocabulary, and even image-style preferences. Point it at the platform where you post most consistently. It's the difference between posts that sound like you and posts that sound like a bot.

9. Know the daily limits before you plan around them

AgentDaily limit
Eva5 files/documents processed
Sonny40 social post drafts
Penny10 blog drafts
Stan200 emails, max 1 every 5 minutes, morning send window

If output stops mid-day, the agent probably hit its cap — it resumes tomorrow. Stan's limits exist to protect your sender reputation, so don't fight them.

10. Test before you automate — especially Stan

Before turning on any schedule or sequence, run one small test task per employee and check the output. And know this about Stan: confirming a suggestion in chat ("start the outreach") can activate a sequence. Keep sequences paused while you're still testing copy and lead filters, and use "Preview with lead" on each step to see the exact email a real lead would get before anything goes live.

Frequently asked questions

Why does my agent claim it can do things it can't?

That's an AI hallucination — every AI tool has them. The fixes: keep Guidelines dialed in, give feedback when it happens, be explicit in instructions, and learn what each agent actually does so you catch drift early.

What should go in the Brain first?

Your pricing, your services, and your brand voice as memories — plus your key web pages. That alone fixes most "generic output" complaints.

How long does setup take?

An honest afternoon to do it right: workspace description, a handful of Brain docs, guidelines per employee, then one test task each. Skipping that afternoon is why most people get mediocre results.

Want your AI team set up right?

Sign up through the link and I'll personally walk you through this exact setup on a free screen-share call — 25% off any plan is applied automatically.

Start with 25% off →

Related: Marblism pricing breakdown · My full Marblism review · Help Center