← Back to all free guides
A free guide by Mariah Brunner ↗ INSTAGRAM ↗ TIKTOK

Train AI Like
A New Hire

Most people quit on AI after one bad answer. The ones who win treat it like a new hire they train over time. Here’s the 3-level setup and the prompts that make it get better every week.

MIT’s NANDA initiative found that about 95% of company AI pilots stall with little to no measurable impact on profit and loss. The cause wasn’t the models. MIT blamed a "learning gap": generic tools don’t learn or adapt to how a company actually works. (Fortune on the MIT report)

I see the same thing with people. They type a vague prompt, get a mediocre answer and quit, like firing a new hire an hour into their first day. The fix is to stop expecting an expert on day one and onboard it over a few weeks.

Onboard It Like An Employee

01

Run the onboarding interview

A new hire knows nothing about your business on day one. You train them on your goals, voice, systems and workflows, and a few weeks in they run real work without you. AI works the same way. Instead of a one-line prompt, let it interview you.

The Onboarding Interview
I want you to work like a trained team member for my business, not a generic chatbot. Your first job is onboarding: learn my business well enough that I could hand you real work. Interview me ONE question at a time and wait for each answer. Cover, in this order: 1. What I do and what I sell (offers, prices, how people buy). 2. Who I serve: my best customers, what they struggle with, the words they use. 3. My brand voice: how I sound, words I love, words I never use. Ask me to paste 2 or 3 examples of my writing. 4. My goals for the next 90 days and my top priority right now. 5. How I like work delivered: length, format, how much detail. 6. The 3 tasks I'd most like off my plate, and how I do each one today. Ask a short follow-up only when an answer is vague. Don't assume anything I haven't told you. When you're done, give me an "Onboarding Brief" with a header for each topic, in short bullets I can paste into my project instructions. End with a list of what you still don't know, so I can fill the gaps.
The 3 Levels, Beginner To Advanced

All that training has to live somewhere, or you start over every chat. Pick the level that fits where you are. You can climb later.

01

Level 1 is a Claude Project

In Claude, click Projects, then New project. Click Set project instructions and paste your Onboarding Brief. Then upload your files, like your brand guide, past work and SOPs. Every chat in that Project starts with all of it. Free accounts get up to five Projects. This is where almost everyone should start.

02

Level 2 is turning repeat work into Skills

Anything you do every week, turn into a Skill Claude runs the same way every time. Open a chat, describe the workflow and attach your templates or examples. Claude asks about your process, writes the skill file and packages it for you. Save it, then make sure it’s switched on in your skills list (under Customize, which is being renamed Plugins). Skills need Code execution and file creation (called Create and edit files in the newer layout) turned on under Settings, then Capabilities. (Anthropic’s walkthrough)

03

Level 3 is Claude Code and agents

When you want to go all in, move into Claude Code. Give it a folder with a CLAUDE.md file that holds your context, and it loads that file at the start of every session. Then build agents that run multi-step work for you. (How CLAUDE.md works)

Make It Critique Itself

01

Run the review at the end of every session

This is the habit that makes everything above compound. Most people get an okay answer, close the tab and start from scratch tomorrow. Instead, have the AI look at what you corrected and rewrite its own instructions. Then paste the new version into your project instructions.

The End-Of-Session Review
Let's do my end-of-session review. Look back over this whole conversation. 1. Misses: where did you miss what I actually wanted? List each time I had to correct you, redirect you or redo something, and what I asked for instead. 2. Lessons: turn each miss into one clear rule you'd follow next time (e.g. "Keep captions under 150 characters," not "be more concise"). 3. Updated instructions: take my current project instructions [paste them here, or say "they're in this project"] and rewrite them with the new rules added. Merge duplicates, cut anything that turned out wrong, and keep it tight. Only include things that would change your output. 4. Changes: list exactly what you added, changed or removed, so I can approve it. Give me the updated instructions in one block I can copy and paste back into my project instructions. Don't invent preferences I didn't show in this chat.

Do this for a couple of weeks and your AI stops acting like a chatbot. It gets better every week because it’s writing its own training from your real feedback.

Don’t Boil The Ocean

Don’t try to train it on your whole business at once. Pick the one task you repeat most, get it running well, then add the next. One workflow at a time is how you end up on the right side of that 95%.