Good morning.
Most entrepreneurs run AI on a scatter of random tasks and then wonder why it hasn't shown up anywhere that matters.
The place it should show up is your P&L.
The difference between AI that feels like a toy and AI that earns its keep is where you point it.
Below are the prompts I'd run first, sorted the way your accountant sorts your business: what adds revenue, what protects margin, and the calls that move both.
—Sam
In Today's Issue
The one-page playbook that makes every prompt sound like you
Five prompts that add to the top line
How to spot the customers ready to buy more
Turn one asset into a week of content
Three margin plays hiding in your workflow
Where you're overpaying for a top-tier model
A weekly review in two minutes, no hedging
Where to start, by business type
Set This Up Once
Before you run any of these, give the LLM a permanent copy of your playbook so you stop re-explaining your business in every chat.
It's a short document with what you do, who you serve, your standards, and how you sound.
Fill in the brackets:
MY PLAYBOOK
WHAT WE DO: [one line]
WHO WE SERVE: [who they are, their size, their situation]
HOW WE WIN: [what makes us different from the obvious competitor]
OUR STANDARDS: [what "good" looks like, and what we never ship]
OUR VOICE: [how we sound, the words we use and the ones we don't]
Now put it where the AI reads it automatically, loaded before you type a word. Pick the one that matches how you work:
Claude (web or Cowork): Create a Project, open its Custom Instructions, and paste the playbook in. Drop your supporting files (your offers, your best past emails, your brand guide) into Project Knowledge. Every chat inside that Project starts with your playbook already in context.
Claude Code, or Cowork on your own folders: Save the playbook as a file named
CLAUDE.mdat the top of your working folder. Claude reads it at the start of every session, no pasting.ChatGPT: Create a Project and paste the playbook into the project instructions, or set it once under Settings, Custom Instructions so it applies to every chat. Add your reference files to the Project too.
Codex: Save the playbook as
AGENTS.mdin the root of your repo. Codex loads it as project rules on every run.
For a one-off chat outside any of these, paste the playbook at the top and go. Either way, every prompt below that says "follow my playbook" reads it first, so the output comes back in your voice instead of a stranger's.
Here's a filled-in scrap so it's concrete, from a freelance packaging designer:
WHAT WE DO: Brand and packaging design for small food and drink makers.
WHO WE SERVE: Founders doing $200k-$2M, scrappy, no in-house designer.
HOW WE WIN: Fast and opinionated, retail-shelf-ready, not endless mockup rounds.
OUR STANDARDS: Three concepts max. Real packaging mockups, never flat logos. Never stock photography. Never miss a Friday update.
OUR VOICE: Warm, direct, a little bold. We say "let's." We never say "synergy."
Two minutes of that, saved where the AI reads it, and every reply stops sounding like a stranger wrote it.
A note on how these are written. Every prompt below leads with the outcome you want, gives the model the context it needs, sets one or two hard rules, and tells it to flag what it's unsure of instead of inventing.
That's what separates a prompt that pays off from a one-liner that hands back mush. Swap the brackets for your details and run them.
Add to the Top Line
These are the prompts that touch revenue directly: leads, replies, upsells, and the content that keeps your pipeline full. Run these first, because they pay for themselves the same week.
Draft outreach that doesn't read like a robot. Personal first messages still convert better than any template, and they're the first thing operators hand to AI and get wrong. Asking for two versions and forcing it to flag thin information gives you a choice without a second prompt and keeps it from making things up about a lead.
Write a first outreach message to [lead]. Goal: earn a reply, not book a call yet. Here's what I know about them: [paste their site, role, a recent post, or whatever put them on my radar]. Rules: under 90 words, no flattery, reference one specific detail that proves I read their world, and end with one low-effort question they can answer in a sentence. Give me two versions with different opening angles. If what I gave you is too thin to personalize honestly, tell me what to go find instead of inventing it. Follow my playbook.
Spot the customers ready to buy more. Your existing list is the cheapest revenue you have, and the buying signals sit in messages nobody has time to read closely. Making it quote the exact line and rate its own confidence means you can trust the flag before you act on it.
Here are recent customer messages: [paste, with names or account IDs]. Find the ones signaling they're ready to buy more: hitting a limit, asking about something a higher tier solves, growing usage, or strong unprompted praise. For each, give me the customer, the exact line that signals it, your confidence (high, medium, low), and a next message that offers help first and the upgrade second. Skip anyone who's only venting. If nobody clearly qualifies, say so rather than reaching.
Turn one asset into a week of content. One good call, post, or script is a week of pipeline if you repurpose it instead of starting from a blank page. Telling it to pull the strongest ideas first, then produce, gets you sharper pieces than asking for the content cold.
Here's a [call transcript / long post / video script]: [paste]. First, pull the three or four strongest ideas in it and list them. Then turn those into a week of content: one short article (400-600 words), three social posts (one hook-led, one story, one contrarian take), and one email that drives to [my offer or link]. Lead every piece with the useful idea, not a warmup. Keep my voice and my playbook. Flag any claim you're not sure I can back up.
See how an AI describes your business to a buyer. More buyers send an assistant to research you before a human ever sees your site. If the assistant can't tell what you do, who it's for, and what it costs, the lead is gone before you're in the room. This shows you what it sees and ranks the gaps by what they cost you.
Here's my homepage text: [paste]. Read it the way a buyer's AI assistant would when they ask it to research me, and answer in plain terms: what do I do, who is it for, what does it cost, and why would someone pick me over the obvious alternative? Then list every question a buyer would have that this page doesn't answer, and the exact lines that read as vague or generic. Rank the gaps by how likely each one is to cost me the lead.
Build a real FAQ from the questions you already get. The questions buyers ask before they commit are the objections costing you sales. Answer them once in your voice, ordered by how close they sit to the decision, and you close the pre-sale gap while giving assistants clean text to quote back to you.
Here are real questions customers ask us before buying: [paste support messages or sales-call notes]. Group them, then write a site-ready FAQ: each question phrased in the customer's own words, each answer plain and under 60 words. Order them by how often they come up and how close to the buying decision they sit. Keep my voice. For any answer you'd have to guess at, write [CHECK] and one line on what you need from me.
Protect the Margin
Revenue you keep is worth more than revenue you chase. These three find money you're already spending and hand it back.
Find where you're overpaying for a model. Running every task on a frontier model is the most common waste I see, because most of what you ask AI to do doesn't need the expensive one. The prompt forces the real question, reasoning versus mechanical work, and hands you a priority list to act on.
Here are the AI tasks I run most, with rough monthly volume for each: [paste]. For each one, tell me whether it genuinely needs a frontier model (real reasoning, nuance, judgment) or would run just as well on a cheaper mid-size or small model (formatting, extraction, sorting, first drafts). Rank them by how much I'm likely overspending, give me a "move this down first" list in priority order, and call out any task where quality would actually drop if I downgraded.
Shrink a bloated workflow. Every extra step is another place the workflow breaks and another line you pay for. Fewer steps means fewer failures and lower cost, so this one separates thinking from information-shuffling, rebuilds it lean, and tells you where it's most likely to drift.
Here's every step I currently run for [task]: [paste the steps]. Sort each step into real thinking (needs a model) or moving information around (a plain tool or a person does it cheaper and more reliably). Then give me the leanest version of this workflow that still hits the same result, name the one or two steps most likely to break or drift over time, and tell me what to check when they do.
Turn a messy process into a checklist. A process stuck in your head costs you every time you're the only one who can run it. Get it out as a checklist a new hire could follow and you can hand it to a cheaper person or an agent, with the risky steps marked so you keep the right ones.
Here's how I do [task], brain-dumped and out of order: [paste]. Turn it into a numbered checklist a new hire could follow without asking me, one action per step. Flag the steps that are risky or easy to get wrong, mark the ones that need a human judgment call, and if any part of my brain-dump is too vague to turn into a clear step, ask me the specific question instead of guessing.
Make the Weekly Calls That Set Your Numbers
Revenue and margin both trace back to the calls you make each week. These make those calls faster and cleaner.
Get a clear decision, not a hedge. AI defaults to a balanced non-answer that helps nobody. Give it your priorities in order and force a single call plus the strongest case against it, and you get a fast second opinion you can act on.
I'm deciding between [options] for [situation]. Here's what matters most to me, in order: [paste your priorities and any hard constraints]. Weigh the trade-offs against those priorities specifically, not in general. Then give me one clear recommendation, the strongest case for the option you're not picking, and the single risk most likely to make you wrong. Don't hedge. If you need one fact to decide, ask for it first.
Run your weekly review in two minutes. Most operators skip the weekly sit-down because it feels like a project. Paste the raw week and let it cut to what moved the business and the one move worth making next.
Here's this week in my business: [paste your key numbers, wins, problems, and anything nagging you]. Give me the three things that actually moved the business, the one thing I'm probably ignoring because it's uncomfortable, and the single highest-leverage move for next week tied to the number it should change. Be direct, cut anything that reads as filler, and ask me for the one number you're missing if it would change your answer.
Turn your week's notes into rules. Keep a running note of the calls you make and why, and this pulls the patterns into directives your team or your AI can follow without you in the room. Do it for a few months and the business runs on your judgment instead of your inbox.
Here are my decision notes from the last few weeks: [paste]. Find the five or six patterns that keep repeating and write each as a short directive my team or my AI could follow without me in the room: a clear "do this" or "avoid this," not a vague principle. Keep my voice, no corporate polish. Mark which ones are strong enough to add straight into my playbook.
Where to Start, by Business Type
Same prompts, but the fastest first move depends on your model.
If you run a SaaS product. Start with the overpaying check; it's the quickest margin in the issue and it compounds every month. Then run the buyer-view audit so people who shop with an assistant find you correctly, and give your in-product AI the playbook so it sounds like your brand instead of a generic bot.
If you run an agency. Productize the "one asset into a week of content" prompt as a service; it's the easiest upsell you have and clients see the output the same day. Run the overpaying check on your own delivery to widen margins, and build each client their own playbook as part of the engagement so the work stays yours.
If you're a freelancer or solo. The overpaying check is pure profit: charge for top-quality output and run the cheaper model on the 80% of work where nobody can tell the difference. Write your one-page playbook today so every deliverable reads senior, and keep your pipeline full with the outreach and content prompts while you're delivering client work.
If you're an expert or creator. Your pipeline runs on content and your voice is the product, so write the playbook first and make it strict about how you sound, because every post and DM inherits it. Lean on the "one asset into a week of content" prompt to turn a single talk, client call, or livestream into a week of material that feeds the audience that feeds your offers. Then run the buying-signal prompt on your DMs and replies to catch the people ready for your next tier before they drift, and use the outreach prompt to warm high-ticket leads without sounding like a funnel.
If you run media or publishing. Point AI at growth (audience, new formats, resurfacing the archive) rather than at trimming the newsroom, which caps your upside. Run the buyer-view audit and the FAQ build so assistants quote you correctly with a link back, then use the content prompt to turn one strong piece into a week across channels.
Last Byte
Each of these is one play you can run today with nothing but a chat window.
Run them and AI models stop being a line item you can't tie to anything and starts showing up where you measure the business.
The prompts are the single play. The system that turns a stack of one-off prompts into agents that run these jobs on a schedule, against a context file that knows your numbers, is what we build in Cortex.
If you want all kinds of full AI systems instead of a single move, that's what's inside.
Talk soon,
Sam Woods
The Editor
.
