Spend Your First Million Muse Tokens

A million tokens sounds like a lot. It isn't. But it's enough to completely change how you work — if you spend them on the right things.

Most people won't. They'll burn through their first million the way tourists burn through foreign currency at an airport: fast, thoughtlessly, on things that don't matter. A poem here. A "write me a business plan" there. Forty-seven clever prompts they found on the internet, each tried once and never again.

Then they'll conclude AI is overhyped, and go back to doing everything the hard way.

Don't be that person. A million tokens is a budget. Spend it like one.

What a million tokens actually is.

Let's get concrete. A token is roughly three-quarters of a word. A million tokens is about 750,000 words — roughly ten novels, or a year of dense email, or about 40,000 back-and-forth conversations with an AI assistant.

That sounds enormous until you realize the model burns tokens on both sides of every exchange. Your instructions cost tokens. Its thinking costs tokens. Its answer costs tokens. A serious working session — the kind where you're actually building something — can run 50,000 to 100,000 tokens without breaking a sweat.

So a million tokens is maybe ten to twenty deep working sessions. That's it. That's your budget.

Which means the question isn't "what cool stuff can I try?" The question is "what do I build first?" Because ten sessions is enough to build real systems, but only if you stop treating tokens like entertainment money.

The mistake everyone makes.

The default way people spend AI tokens looks like this: they open the chat, type something clever, get something impressive back, and feel productive. Repeat until the tokens are gone. Nothing in their life is different.

I call this prompt collecting. It's the AI equivalent of buying gym equipment. The prompts are great. The collection is impressive. Nothing changed.

Here's the uncomfortable truth: a single great answer from an AI is worth almost nothing. It's a sugar rush. What compounds is a system that gives you good answers every morning without you having to think of the question.

The people getting real value from AI aren't the ones with the best prompts. They're the ones who stopped asking questions and started building infrastructure. They spent their tokens on systems, not moments.

That's what this guide is. Not ten prompts to try. A spending plan.

Spend one: your operating manual.

The highest-leverage thing you can do with your first chunk of tokens is write down how you work — and teach it to the machine.

I mean this literally. Write a document that describes: what you do, what you care about, how you make decisions, what "urgent" means to you, what you never want to see, how you like information presented. The stuff that's obvious to you and invisible to everyone else.

Then feed it to Muse as a standing instruction. Not a prompt you paste each time — a persistent system prompt, a briefing configuration, the thing it reads before it does anything for you.

This will feel tedious. It will feel like it shouldn't matter. Do it anyway.

Here's why: every future interaction gets better. Every morning briefing, every inbox triage, every research task — all of it runs through the lens of who you actually are instead of a generic helpful assistant guessing at what you want. The difference between "here's the news" and "here's what matters to you in the news, ranked by what you're actually working on" is this document.

Budget: 50,000–100,000 tokens. Most of it is the back-and-forth of refining it. You'll write a draft, test it, correct it. "No, that's not what I meant by urgent." "Yes, always flag that." "Never show me that format again." Each correction makes every future session better. This is the highest ROI spend on the list.

Spend two: automate one painful thing.

Pick the single most annoying recurring task in your week. Not the most impressive one. The most annoying one.

For me it was the morning. I'd wake up and spend twenty minutes figuring out what day it was — checking the calendar, scanning overnight news, trying to remember what was actually due versus what just felt due. Twenty minutes, every day, of pure cognitive overhead before I'd done anything useful.

So I built a morning briefing. Muse pulls overnight news in my field, reads my calendar, looks at my task list, and hands me a ninety-second read: here's today, here's what matters, here's the one thing to focus on first. It runs on a schedule. I don't ask for it. It just shows up.

Twenty minutes a day is over 120 hours a year. I spent maybe three working sessions building the briefing. The math isn't close.

Your painful thing is different. Maybe it's inbox triage. Maybe it's meeting prep. Maybe it's the weekly report nobody reads but everybody requires. Whatever it is — the thing you dread, the thing you procrastinate, the thing that drains you before the real work starts — that's your second spend.

Don't try to automate five things. Automate one. Get it working. Feel the difference. Then decide if you want a second.

Budget: 100,000–200,000 tokens. You'll iterate. The first version will be wrong in interesting ways. That's fine — the corrections are the product.

Spend three: build a research workflow.

Here's a pattern I use constantly: before any significant decision, I have Muse do a deep research pass. Not "tell me about X" — a structured investigation with explicit confidence grading.

The workflow looks like this: I give it the question and the context. It goes and reads — articles, docs, threads, whatever's relevant. Then it comes back with findings, and critically, it tells me where it's confident and where it's guessing. "High confidence on the pricing data, medium on the timeline, low on the competitive claims — those come from a single source."

That last part matters more than everything else. An AI that tells you what it doesn't know is worth ten that confidently hallucinate. But it won't do this by default. You have to build it into the workflow: explicit instructions to grade confidence, cite sources, flag gaps.

Once this workflow exists, you use it for everything. Job decisions. Purchase decisions. "Is this legit or a scam." Technical architecture choices. It becomes the thing you run before you commit to anything significant.

Budget: 50,000–100,000 tokens to build and refine. Then it costs a few thousand per use, forever.

Spend four: set up the things that run while you sleep.

This is the spend most people never think of, and it's the one that feels most like magic once it's working.

Muse can run scheduled tasks. Not "remind me to do X" — actually do X. Every morning at 6:30, check these three sources and send me anything new. Every Monday, review my open projects and flag what's stalled. Every evening, draft tomorrow's priority list based on what's due.

The key insight: you're not just saving the time it takes to do these things. You're saving the mental RAM of remembering to do them. The number of things I used to carry in my head — "check if that deployed," "follow up on that thread," "see if there's news on that topic" — that now just happen without me thinking about them is the real win.

Start with one scheduled task. Something you'd check manually every day if you had infinite discipline. You don't have infinite discipline. The machine does.

Budget: 50,000 tokens to set up your first few. The ongoing cost is trivial — these are small, focused runs.

Spend five: the boring correction loop.

This is the spend nobody wants to hear about, and it's the most important one.

Every system I described above got good through correction. Not clever engineering — boring, repetitive correction. "That's wrong." "Closer, but not quite." "Yes, exactly like that, always do it that way." Dozens of small corrections, compounded.

People want the AI to be smart out of the box. It's not. It's a very fast, very knowledgeable intern with no context on your life. The context is the product, and you build it one correction at a time.

Here's what this looks like in practice: you run your morning briefing. It's 70% right. You correct the 30%. Next day it's 75% right. You correct the rest. After two weeks it's 90% right and you barely look at it before trusting it. After a month it's part of your life.

Nobody posts about this on social media because it's not impressive. It's just... work. Unsexy, repetitive, compounding work. The same kind of work that makes everything else in life actually function.

Budget: ongoing. A few thousand tokens a day in corrections. This never stops, and that's the point — the system keeps getting better as your life changes.

What not to spend on.

Don't spend your first million on party tricks. The poem, the "roast my resume," the "write a story about a pirate accountant" — fun for five minutes, worth nothing after.

Don't spend it collecting prompts from the internet. Someone else's prompt is someone else's system. It won't know your calendar, your priorities, your voice, your life. The prompt isn't the value. The hundred corrections are.

Don't spend it trying to replace judgment. AI is excellent at laying out options, finding blind spots, doing the research. It's terrible at making the call. Anyone outsourcing decisions to a model is going to learn an expensive lesson about accountability. The call is yours. It always was.

And don't spend it on things that need to sound like you. I write my own blog posts. I write my own important messages. The uncanny almost-right version of your voice is worse than just doing it yourself. Use AI for the scaffolding, not the soul.

The actual budget.

Here's how I'd allocate the million, roughly:

• Operating manual: 100,000 tokens
• First automation: 150,000 tokens
• Research workflow: 75,000 tokens
• Scheduled tasks: 50,000 tokens
• Correction loop (first month): 100,000 tokens
• Buffer for experiments and mistakes: 525,000 tokens

Yes, half the budget is buffer. You'll waste tokens. You'll build something that doesn't work. You'll go down a rabbit hole. That's fine — that's learning. The plan isn't to spend every token optimally. The plan is to spend enough tokens on systems that the rest doesn't matter.

After the first million, you'll know exactly what the second million is for. That's how you know the first one worked.

The real answer.

Someone's going to read this and ask for the prompts. The exact wording. The templates. The system prompt I use for my briefing.

I'll save you the trouble: they won't help you. My briefing prompt is the product of months of corrections on my life. Your life is different. Your priorities are different. What "urgent" means to you is different.

The prompts aren't the thing. The process is the thing. Sit down, write out how you work, build one system, correct it until it's good, then build the next one. That's the whole guide. Everything else is details.

A million tokens is enough to change how you operate. Most people will spend it on entertainment and conclude AI is a toy.

Don't be most people. Spend it on infrastructure. The toy phase ends. The systems compound.

That's what I'd do with my first million. It's what I did with mine.