How To Train Your AI Assistant

Before you can manage an AI employee, you have to stop using AI like a search engine. We open with a direct question: what does AI actually look like to you? Most people's mental model is some combination of magic box, Google, and intern and that confusion is exactly why their results are mediocre. We dismantle that picture and install a new one: the employee model that everything else in the course hangs from.

Then we name them. Every participant leaves this session with at least one named, versioned AI employee (e.g. Samantha v.3, Copywriter) and a clear sense of what category of work they're hiring for. The naming sounds light but it's load-bearing. It's what lets you treat the AI as a colleague, track its versions, fire it, retrain it, replace it.

We close with a frank conversation about tools: which ones are worth your time right now, what "agent" actually means and when to use one, and the basics of using AI responsibly with sensitive or personal data.

You wouldn't hire a human without a job description. Most people hire AIs without one every day, and then wonder why the output is generic. This session is about writing the role profile that makes the difference between a useful AI employee and a frustrating one.

You'll work through the five-part anatomy of a role profile: Mission and Scope, Responsibilities with specific example tasks, Limits, Inputs and Expected Outputs, and KPIs. The Limits section is where most people skip and most failures happen — what is this employee explicitly not allowed to do, and what requires your sign-off? We'll write both, prompt-side and policy-side.

You'll leave this session having drafted a complete role profile for one AI employee in your own work. Bring a real use case. We'll use a shared template you can reuse for every subsequent hire.

Once you have a role profile, you need to know whether the AI is actually doing the job and whether the job is worth doing this way at all. We start with the honest question every AI project has to answer: is this actually saving time? Two rules govern this. Never sacrifice quality. And if the AI takes as long or longer than doing it yourself, don't use it.

• Then we build the workflow around your role profile.
• When do reviews happen?
• What KPIs matter?
• When do you step back in?
• When do you spawn a new version of an existing employee versus hiring a new one entirely.

These questions sound operational because they are, and they're where most projects quietly collapse.

We'll also work through identity and voice creation. Pick a character, archetype or person your AI should sound like, then stress-test it with two prompts, one creative, one charged to see whether the voice holds under pressure.

Most people think "training" an AI means writing better prompts. Prompts are one layer of five. This session covers the full stack — Data → Prompts → Tools → Policies → Voice, and you'll leave with a clean training pack for one AI employee.

We go deep on each layer. Data: what your AI needs to know, how to assemble it cleanly, and the legal and privacy considerations that quietly sink projects. Prompts: moving beyond casual prompting into prompt architecture, with two real specimens dissected live. Tools: when your employee should call a calendar, email, code, or a knowledge base, and how to set permissions like you would for any new hire.

Then Policies and Voice: the layers that make an AI employee trustworthy. Policies cover the never-do list, escalation rules, and automatic guardrails. Voice is what stops your AI sounding generic. We close with the most important rule in the entire course: the Final Voice rule. Anything intended to have emotional impact on a real human has to be read, edited, and signed off by a real human. AI drafts. Humans publish.

By Week 5 you've designed an AI employee on paper and trained it on the stack. This session is where it gets bones. No matter what your technical background is - and most participants on this course are not coders - you'll walk away with a fully coded mini app that Claude builds with you, live, in Claude Code.

We start by translating one of your role profiles into something Claude Code can actually run, then we build it together. You drive, Claude Code executes, Isvari coaches. By the end of the session you'll have a working mini app, small, functional, yours, that does a real piece of work for you. Not a demo. Not a tutorial. The thing.

We close with what changes when your AI employee can write and run code on your behalf: new permissions to set, new audit trails to keep, new questions about what it should and shouldn't be allowed to touch.

Week 5 gave your AI employee hands. Week 6 gives it a desk, a calendar, an inbox, and a to-do list. We use Claude Cowork to build an AI assistant that can handle real tasks for you autonomously overnight, between meetings, across the gaps in your day where work currently piles up.

We start by mapping where time leaks in your week: inbox triage, scheduling, follow-ups, the same five messages you write in slightly different forms and choose two or three tasks to delegate. Then we build the assistant together: integrations, operating instructions, what it should do unprompted, what it should draft and wait for your sign-off on, what it must never touch.

We close with the management discipline this kind of AI demands. An AI that acts needs different oversight from an AI that suggests. We'll set up the audit log, the daily review habit, and the kill switch you should have ready before you ever let an AI send something on your behalf.

Hiring is the easy part. Managing performance over time is what separates people who get lasting value from AI from people who just paid for subscriptions. We start with the four levels of human review, L0 (no oversight) through to L3 (complex feedback on edge cases) and how to choose the right level for each AI employee.

Then the performance management loop: Test → Observe → Error Taxonomy → Fix → Retest, with a real red-team exercise you can run yourself.

Then we cover the three main collaboration patterns: Draft & Edit (you as user), Collaborator (AI as a peer), and Orchestrator (AI coordinating other AIs). Each fits different work and different risk profiles. We'll also cover what humans do best; judgement under ambiguity, emotional weight, genuine creativity and how to design your workflow so you keep those moments and your AI gets the rest.

We close with retraining and versioning: naming conventions, change logs, rollback plans. The goal is to be able to answer "what did this AI do differently last month?" and to roll back when an update makes things worse.

The last content session is about turning everything you've built into a routine that actually sticks, anchored on a 30/60/90 day plan.

First 30 days: finalise your role profiles, choose your tools, build your system prompts and voice packs, and start using your AI employees for real. Days 30–60: review what's working, decide which AI employees get retained, promoted or retired, and run a red-team week. Day 90 and beyond: monthly reviews, honest feedback loops, and formalised habits that keep your AI workforce sharp and accountable.

We'll work through a weekly operating cadence a 30-minute weekly review template you can adopt as-is and close with the four pitfalls to avoid: prompt roulette, data dumps, no ownership, and flying blind on results. And the question you should always be able to answer: when do you retire an AI employee?

This is where the course gets specific to your situation. First, present your favourite AI employee to the whole group. Then Isvari fields the real questions you have about what comes next.

Come with: an AI employee you've designed and want feedback on; a project that's stalling and you can't diagnose; a compliance, legal or ethics question you're wrestling with; a specific tool, prompt or framework that didn't translate to your context; or anything that didn't make sense - or made too much sense and needs interrogating.

This session is deliberately unstructured. It's where the most valuable learning tends to happen, because the questions are real, the projects are live, and Isvari's answers are honest.