Everyone around me has been switching to Claude. I did too, and I went deep. Projects, context documents, voice profiles, data uploads. I genuinely thought my setup was advanced. Then I sat down with Allie K. Miller and she showed me what advanced actually looks like.
Allie built IBM's first multimodal AI team, ran machine learning for startups at Amazon, and is on the TIME100 AI list. Today she runs Open Machine, an AI advisory firm working with Google, Samsung, Novartis, Warner Bros.
Allie has about 100 AI agents running while she sleeps. Every morning she wakes up to a full briefing — industry news, calendar prep, drafted email replies ranked by urgency. Zero code.
When Allie told me that most people are using AI at 20% of its potential, I thought: okay, not me though.
Then she walked me through her actual setup and I realized — I was maybe at 30%. Good setup. Real output. And still so much room.
Here are 5 things most people get wrong when starting using Claude.
1. Using Claude as a search engine instead of a team member
Most people open Claude and ask it questions. "What should I write about?" "Give me 10 ideas for my newsletter." They get back 10 perfectly generic answers that could've come from any AI blog.
The fix is giving Claude your actual data and assigning it a job. I upload my newsletter performance CSVs, subscriber acquisition reports, and top URLs — then say: "You're my newsletter strategist. Based on this data, which topic has the highest open rate potential for next Tuesday?"
Completely different output. Specific to my audience, backed by real numbers.
Allie calls this the shift from "microtasker" to "delegate."
I get pretty annoyed when people say AI is an intern. What intern has PhD-level intelligence and can read the entire internet?
2. Starting every conversation from scratch
This is the one I see everywhere. People re-explain their tone, their business, their preferences every single session. Copy-paste the same context. Sometimes forget something and the output is completely off.
The fix: a Claude Project with everything inside. I have my voice profile (how I sound when I'm excited, what phrases I'd never use), a fact sheet about my company and podcast guests, an anti-AI writing checklist, and 6 months of performance data. Claude already knows me when I open it.
Allie recommends starting with three "context documents": a personal constitution (your values, how you think), a goals doc (specific enough for AI to reference), and a business strategy doc (including what failed and why).
Her team built all three in a single one-hour Zoom — cameras on, everyone on mute, just answering Claude's questions. She called it a "context hack".
3. Publishing the first draft
This one's a trap. Claude writes fast and the first version sounds... fine. Polished, even. So it's tempting to make a few edits and hit publish.
I learned to compare my AI-assisted drafts to my best posts from before I used AI. The difference was obvious.
AI drafts had the same rhythm in every paragraph. Same constructions repeated throughout. Tidy conclusions that restated everything I'd just said. One draft had 6 "rule of three" patterns in 800 words. (You know it: "It's not about speed. It's not about volume. It's about consistency." That one. Six times.)
My whole team now uses an Anti-AI Writing Guide that flags these patterns across every project. Now Claude catches its own habits before I have to.
That guide turned into one of my most downloaded documents. Turns out, everyone has this problem.
4. Using Claude for things it's bad at
My content producer once uploaded a 19-minute audio recording for transcription. Claude tried 5 different methods over 40 minutes. Each one failed. It finally told me it couldn't do it. A tool like Otter.ai would've taken 2 minutes.
Quick list of what Claude is not best at so you don't waste time: audio transcription, generating images, and giving you real-time numbers you'd bet money on.
Allie also made an important point here. Most people use the basic web chat for everything — but Claude actually has four versions:
Claude Chat | Questions and research |
Claude Cowork | Can touch your files and take action |
Claude Code | Full control, builds software |
Claude Chrome extension | Controls your browser |
Using version 1 for everything is like owning a toolbox and only using the hammer.
5. Writing perfect prompts instead of just complaining
This one comes directly from Allie and it's the weirdest advice I've gotten this year.
The best way to write a prompt is to just complain.
Not "please create a workflow for my morning briefing with the following specifications." Instead: "I feel really stressed every morning. I never know what's happening in my industry before meetings. I keep getting rained on. And I have no social life because I never know what events are happening in my city."
She did this live during our episode. Rambled for about 60 seconds. Claude came back with a full system: industry news ranked by what would impress her boss, weather and outfit suggestions, local events for the next 4 days. No code. About 3 minutes total.
I've been trying this for two weeks and it works better than my structured prompts.
When you complain, you accidentally include all the context that matters — the real problem, the specific situation, the emotional weight. When you write a "professional" prompt, you strip all that out and get a sanitized answer.
If I were starting today
Two things, day one.
60 minutes building context documents. Voice profile, business facts, goals. Upload to a Claude Project.
Stop prompting. Start complaining.
The episode with Allie is one of the most practical episodes I've recorded. She builds a morning briefing agent live from scratch, walks through her 36-workflow system, and explains the 4 levels most people never get past.
If you're stuck on something with Claude — a workflow that isn't clicking, a task you can't get right — reply and tell me. Best questions to become next editions.
Talk soon,
— Marina 💜