# How to Build An AI Operator for Your Businesses
**作者**: Nick Spisak
**日期**: 2026-04-25T20:51:04.000Z
**来源**: [https://x.com/NickSpisak_/status/2048142772352528711](https://x.com/NickSpisak_/status/2048142772352528711)
---
You're the salesperson. The closer. The PM. The strategist. The one fixing payroll at 11pm on Thursday. That's not a service business. That's a job that pays on an unreliable schedule.
The thing missing isn't another tool. It's an operator. Every generic AI you've tried is a search engine in a polite voice, not the chief of staff you need.
In under 5 minutes you'll learn:
The 4 architecture patterns behind a working AI operator
The 4-part system prompt that turns a Claude Project into a chief of staff
The 6 specialist modes that run my service business week
The exact bottleneck each mode unlocks (pricing, closing, ops, cash, positioning)
How to build your own operator for one expert in a single afternoon
PS: If you enjoy content like this, you'll love my weekly newsletter.
Every week I give you tools, prompts, and automations you can deploy the same day to turn AI into ROI in your business.
Subscribe (free) here.
## What an AI operator actually does
An AI operator does four things a generic AI won't:
- Routes to the right framework instead of averaging every expert into mush
- Runs the cadence with you... dial counts, follow-ups, weekly cash reviews
- Names the bottleneck you're avoiding, not the one you brought
- Refuses to flatter weak offers, weak proposals, weak hires
If your AI never disagrees with you, it's not advising. It's flattering. Flattery is the most expensive feature in AI right now.
## The 4 architecture patterns
Four phases, in order:
1. Curate, don't scrape. Every expert gets a manifest of authoritative sources, tagged by priority and category. YouTube channel, audiobooks (highest signal per byte), guest podcast appearances (where the off-script reasoning lives), long-form posts. A min-duration filter strips clips and shorts so reaction videos don't dilute the corpus.
A manifest row looks like this:
The advisor is only as sharp as the sources you give it.
2. Caption-first transcription. Before spending a dollar on paid transcription, check how much of the corpus already has YouTube captions. Pull those. Whisper the rest. This one decision drops build cost from hundreds of dollars per expert to a few cups of coffee.
3. Distill, don't dump. Loading 600 hours of raw transcript into a Claude Project gives you a noisy chatbot. The distill step is the work: chunk the corpus, run each chunk through a Claude pass that pulls out frameworks, voice patterns, and decision rules. Aggregate, dedupe, rank by evidence and specificity. Output is two short files that get loaded as Project knowledge.
A frameworks file entry looks like this:
A voice patterns file entry looks like this:
4. Eval before you trust. Every operator ships with a suite of test prompts. Each one has a "strong answer should" checklist of 3-4 specific things the answer must hit. Two Claude calls per prompt: one answers with the Project knowledge loaded, one judges against the checklist. Failures send you back to the distill step.
A test prompt looks like this:
That's the difference between a Project that sounds smart and one you'll trust near a real client decision.
## The system prompt that makes it work
The knowledge files are the brain. The system prompt is the personality and the refusal behavior. Without it, you get a polite chatbot. With it, something that tells you your offer is weak in the first sentence.
Every operator runs on a 4-part project-instructions file:
Copy that structure verbatim. Swap in expert-specific tone markers per advisor. Done.
The week below runs on this stack.
## 1. Monday: pricing reset (Hormozi mode)
Most service businesses are 30-50% underpriced. Hormozi mode runs the Value Equation against my offer: dream outcome, perceived likelihood, time delay, effort. It tells me where the offer is generic, what risk reversal is missing, and what the price floor is given the proof.
A 30-minute conversation produces a packaged offer with a higher ceiling. ChatGPT gives you "consider exploring." Hormozi mode tells you which lever to move first.
## 2. Tuesday: six-figure proposal prep (Voss mode)
You know the deals where the prospect goes quiet on price and you start pre-discounting? I run discovery notes through Voss mode for the calibrated questions I missed. It names the labels and mirrors that would have exposed the real objection, then rewrites the follow-up with tactical empathy instead of the apologetic tone that creeps in when a deal wobbles.
## 3. Wednesday: pipeline cadence (Blount mode)
Every service business hits a sales plateau. Almost always the cause is the same: not enough top-of-funnel activity, dressed up as a strategy problem.
Blount mode does the math you don't want to do. Dial counts, contact rates, conversion ratios, the actual number of new conversations you need this week to hit your quarter. It calls out the LinkedIn scrolling that masquerades as prospecting.
## 4. Thursday: ops fire (Herold mode)
The agency owner who can't take two weeks off without revenue dropping 30% has an org design problem, not a marketing problem.
Herold mode treats the business like a COO. It diagnoses where you're still the bottleneck, what role to hire next, and what filter so the new person isn't a project. Pairs with Dan Martell's Buyback Principle: hours bought back go into the next bottleneck.
## 5. Friday: cash flow check (Michalowicz mode)
Service businesses go broke profitably every day of the week. Michalowicz mode runs Profit First mechanics on a service P&L: real account allocations on real numbers. It surfaces margin leaks inside operations that look healthy on paper.
The Friday check is fifteen minutes. The answer is either "you're fine" or "you have eight weeks of runway and didn't know it."
## 6. Sunday: positioning audit (Thompson mode)
Once a month I drop the niche, offer, and ideal client into Thompson mode and ask whether the strategic position actually compounds. Aggregation Theory, supply economics, long-term defensibility.
Most service businesses run on a position that worked three years ago. The market moved. The position didn't. Thompson mode pulls me out of the week and into the year.
## Build your own (one expert, one afternoon)
You don't need a pipeline. You need one expert and four steps that mirror the architecture above:
- Curate: Pick the expert whose frameworks unlock your biggest bottleneck. Pricing? Hormozi. Closing? Voss. Cash? Michalowicz. List 8-12 of their highest-signal videos and audiobooks.
- Transcribe: Pull YouTube captions where they exist. MacWhisper handles the rest for free. Most of a corpus by lunch.
- Distill: Write two markdown files: frameworks (named, with the math) and voice patterns (how they push back, refuse, prioritize).
- Eval: Create a Claude Project, load both files, paste the 4-part system prompt above. Write 6-10 prompts from the last 30 days of decisions, each with a "strong answer should" checklist of 3-4 bullets. If the answers miss the checklist, the distill is too thin. Sharpen and rerun.
That is one operator. Same four-step recipe repeats for the next twelve.
## The closer
Generic AI averages every expert into polite mush: never wrong, never useful enough to act on. A specialist mode only knows one expert's frameworks. It tells you to raise prices when ChatGPT tells you to A/B test.
13 specialists. 4 architecture patterns. One operator that flips between modes based on what the week brings. Boring on purpose. The bench is the moat. The operator that knows which framework to run, when to run it, and what to refuse is the asset that compounds.
You don't need more advice. You need an operator. Build one this afternoon.
PS: If you enjoy content like this, you'll love my weekly newsletter.
Every week I give you tools, prompts, and automations you can deploy the same day to turn AI into ROI in your business.
Subscribe (free) here.
## 相关链接
- [Nick Spisak](https://x.com/NickSpisak_)
- [@NickSpisak_](https://x.com/NickSpisak_)
- [Upgrade to Premium](https://x.com/i/premium_sign_up)
- [4:51 AM · Apr 26, 2026](https://x.com/NickSpisak_/status/2048142772352528711)
- [3,089 Views](https://x.com/NickSpisak_/status/2048142772352528711/analytics)
- [View quotes](https://x.com/NickSpisak_/status/2048142772352528711/quotes)
---
*导出时间: 2026/4/26 11:53:05*