The Rye Prompt Stack | 11 Ways I Use Frontier Models on Real Problems

Why I Think These Work

Greg built these around a simple idea: most people prompt like they are filling out a form. They give the model a task, it gives them an answer, they move on. That is fine for simple stuff. But for real problems, you need the model to hold your whole situation in its head and then argue with itself about the best answer.

The real leverage is not in the prompt. It is in the context you give before the prompt and the permission you give the model to disagree with you.

Here is what makes these prompts different from most of what you will find online:

It Holds the Whole Picture

Drop in two years of decision docs and tell it to find your blind spots. No spoon-feeding. No summarizing. Just give it everything.

It Can Argue With Itself

Assign it conflicting personas and make them score your work against each other. The best answer wins, not the most agreeable one.

It Finds What Is Missing

It has seen enough successful businesses to know what yours should have. Ask it to look for gaps, not flaws.

The Core Eight

Prompts I Run When I Need Real Answers

The Sparring Partner

Before you build anything, make the model find the holes in your idea. One question at a time.

I want to build this. Do not start building yet. Interview me the way a founder who has shipped three successful products would. One question at a time, 15 max, hunting for the thing I have not thought about. Push back when my answers are vague or optimistic. After the interview, write the full spec, list three ways this fails, and then build the V1. [Attach your rough idea]

Ryan's take: The "list three ways this fails" instruction is the most important part. Most people skip it. That is where you find the real problems before you spend 40 hours building something nobody wants.

The Blind Spot Sweep

Do not ask what is wrong with what you have. Ask what is completely missing.

Here is my analytics, my content calendar, and my current offer. Do not critique what exists. Spend your time on what is absent. What is every successful company in my space doing that appears nowhere in my business? What customer segment am I not even trying to reach? What revenue line should exist and does not? Rank everything by estimated money left on the table.

Ryan's take: I run this every quarter. The "rank by money left on the table" instruction is what makes it useful. Without it you get a brainstorm. With it you get a prioritized action list.

The Copy Cage Match

Eight drafts, five judges, one winner. Stop settling for the first thing it writes.

Write my landing page eight different ways. Different hook, different structure, different angle each time. Then assign five judges: a skeptical CFO who controls the budget, a founder scrolling at midnight who is half asleep, a direct competitor who wants to undercut me, my ideal customer who has been burned before, and a direct-response copywriter who has seen everything. Have every judge score every version with a number and a reason. Kill the losers. Merge what worked into one final version and show me the full scoreboard. [Attach your product or startup idea]

Ryan's take: The "founder scrolling at midnight who is half asleep" judge is the one that saves your copy. That is your actual reader. If your hook does not survive that persona, it does not go live.

The Hostile Takeover

Give it everything and tell it to destroy you. Find your vulnerabilities before a competitor does.

Here is my P&L, my pricing page, my churn data, and my last 50 support tickets. You are a well-funded founder who has studied my business and wants to take my customers. Build the company that replaces mine: positioning, pricing, the first 10 customers you would steal, and the exact outreach you would send each one. Then rank every attack by how much damage it could do with minimal resources. Do not soften this. I need to hear it clearly.

Ryan's take: The "rank by damage with minimal resources" line is what makes this actionable. It turns a threat analysis into a real priority list. The things that score highest are the ones to fix first.

The Pattern Mirror

Feed it two years of your decisions and let it find the patterns you cannot see from the inside.

Here are two years of my notes, decision logs, and post-mortems. Map every meaningful decision I made: what I chose, what I believed at the time, and what actually happened. Find the patterns. Where am I systematically too early, too late, or too optimistic? What do I always say right before a bad call? Write me the one-page operating manual for working with me that a good COO would quietly keep in their back pocket.

Ryan's take: No old notes? Use your emails and messages. The pattern is in there. And once you see it, you cannot unsee it. This is one of the most useful things I have ever run.

The Deal Table

Put a real negotiator across from you before the actual conversation happens.

I am about to negotiate a deal with [counterparty]. Here is everything I know about them, what they want, and what I want. Become them. Their incentives, their alternatives, their pressure points. We go round by round. You respond the way they actually would, including going quiet, getting aggressive, or asking for time. After each round, step out of character and tell me what I just gave away for free. Do not let me win because I am the one typing.

Ryan's take: "Do not let me win because I am the one typing" is the key line. Without it the model will be polite and let you off the hook. You need it to be honest, not agreeable.

The Fine Print Finder

The buried costs are always in the tables, footnotes, and schedules. Use this before your lawyer review.

Here is the full contract including every exhibit and schedule. Read all of it, especially the tables. Give me three lists: what costs me money that is not obvious on first read, what I would regret in 18 months, and what is missing that should be here to protect me. Then give me the three changes I should request and exactly how to phrase each ask. Flag anything where I should get a real lawyer involved before signing.

Ryan's take: This is not a lawyer replacement. Use it to find the flags before your legal review so you are not paying by the hour for things you could have spotted yourself.

The Self-Builder

After a few weeks of use, make it study its own work and build tools for the things you keep asking.

Look back at everything I have asked you to do over the past month. Find the requests that repeat or follow the same pattern. Build reusable instructions and tools for each one so next time it takes one sentence instead of ten. Then tell me: based on what I keep asking, what should I be delegating to you that I am still doing manually?

Ryan's take: This is the prompt that separates people who get it from people who do not. If you are not building reusable systems out of your AI interactions, you are starting from zero every single time.

The Business Models

3 Prompts That Are Also Businesses

Each of these is a prompt you can run for yourself and a service you can sell. The business model is built into the prompt design.

The Persona Panel

Build customer personas from real review language, not demographics. Then score ad variants overnight.

Here are 500 customer reviews for [brand] and their 10 strongest performing ads. Build six customer personas using the actual language from the reviews, not demographic categories. Then take these new ad concepts and have every persona evaluate each one: Would they stop scrolling? Would they click? What is their specific objection? Output a kill list, a launch list, and the one insight about this customer that the brand is clearly missing. [Attach reviews and ad concepts]

Ryan's take: "Build personas from actual review language, not demographics" is the instruction that makes this work. Demographics are guesses. Review language is what customers say when they think nobody is watching. That is the real signal.

The 48-Hour Build

Scope a custom internal tool live on a call. The interview is the sales demo.

You are about to interview a small business owner about the most painful repeated process in their operation. One question at a time, 12 questions max, plain English only, no technical terms. Dig until you can describe their workflow better than they can. Then write the spec for the smallest possible tool that removes the pain, build it, and write a one-page guide that their least technical employee can follow without any help.

Ryan's take: Scoping was always the expensive part of custom software. This prompt does the scoping live in front of the client. They watch the spec being built in real time. That is your discovery call and your demo in one conversation.

The Contract Sweep

Read the vendor contracts nobody has opened since signing. Free unless you find money.

Here are all of our vendor contracts and 12 months of invoices. Cross-reference them carefully. Find auto-renewals coming up in the next 90 days, invoices that do not match the contracted rates, seats or licenses we are paying for but not using, and price escalators we have the right to challenge. Rank everything by dollar impact. For the top 10 findings, draft the exact email to send the vendor including the leverage we have and the specific number to ask for.

Ryan's take: Every company with more than 10 vendors has contracts nobody has opened since signing. The offer is simple: free unless I find money, then I keep a percentage of what I save you. The model does the reading. You do the one phone call.

11 Prompts Worth Knowing About (Credit: Greg Isenberg)