A prompt that actually works isn't a vending-machine code, it's a conversation, and it stacks six elements: Role (who the AI is), Context (who you're talking to plus who you are), Task (the action and the why), Format (the shape you want), Constraints (the guardrails, especially the negatives), and Examples (what right looks like). Pile those six into one prompt and the output stops sounding like the internet talking to the internet and starts sounding like you. Then you wrap three conversation patterns around it (the Interview Flip, the Redirect Loop, and the Mirror) so you never accept version one. Build that prompt once, hand it to your teammates, and the whole team gets consistent output. That's not a prompt anymore. That's a system.
That's the build we walked through in Session 4 of our free 8-part series, Building Your AI Content System. The full 33-minute replay is up top. Hit play, or keep reading for the recap.
Why Do Most AI Prompts Fail?
Most AI prompts fail because of a communication breakdown between the human and the AI, not because the model is dumb.
Let's be honest about that, because it's the reframe that changes everything. The models just got really, really good. ChatGPT shipped 5.5 a couple weeks ago. Anthropic released Opus 4.8 yesterday at four in the afternoon. The tools are sharp. So when a prompt falls flat, the problem usually isn't the machine. It's how the human asked.
George has seen this exact pattern his whole career, and not just with AI. Inside organizations, when things break down between humans, it's almost always a conversational breakdown. Same thing here. The fun part: humans are hard to fix. AI is easy, because AI is just structure.
It's not that the tool isn't good. It's that there's a communication breakdown between the human and the AI assistant.
Here's the trap most of us fall into. We treat the prompt like a vending machine. Punch in a little code, expect magic to drop out. "Write me a blog post." That's the whole prompt, and it's terrible. What comes out tastes like a snack that's been sitting in the machine for six years. Cardboard.
So before you hit enter, look at the thing you typed or spoke and ask one question:
How many guesses are left in this prompt before I hit enter?
The more guesses you leave in there, the more garbage you get back. Garbage in, garbage out, we all know that one. And what we're really fighting against is generic context. That's all "bad" or "broken" actually means: generic context. The good news is it's easy to fix.
At Sidekick Strategies, we spent years making HubSpot work for humans. This is the same mindset pointed at a new target: making AI work for you and the humans you serve.
What Are The Six Elements Of A Great Prompt?
The six elements of a great prompt are Role, Context, Task, Format, Constraints, and Examples. Stack as many of them as you can into every conversation and the output gets dramatically better. You won't need all six every single time, but the more you load in, the sharper the result.
Let's walk each one with George's specifics.
Role: Cast An Actor
You tell the AI who it is. Think of it like casting an actor in Hollywood. "You are a seasoned email marketing strategist." "You are a friendly customer success coach." That one line sets the perspective for everything that follows.
It takes five seconds to write. The payoff is huge, because now the AI is answering from a specific point of view instead of from the bland average of everything it knows.
Here's the level-up: you can chain roles. For the first couple years of his AI journey, George would write something like "you are a content strategist who's well versed as a writer and has also done some editing." Read that back. He just built a writer, a content strategist, and an editor into one assistant. One role or three. Your call.
Context: Write To Bob, Not Everyone
This is the highest-impact element of all six. Context is who you're talking to, what situation they're in, and what's keeping them up at night. The audience, the scenario, and the pain point.
It might sound like this: "My audience is solo consultants who are overwhelmed by tech and short on time." Simple line, three elements packed inside. The more your AI knows about the human on the other end, the more it writes for that human instead of for everyone.
Write to Bob. Write to Jenny. Not everyone.
If you improve only one thing after this session, make it context. Audience, scenario, pain point.
And here's the flip side of the coin. So far we've talked about context about your reader. There's another kind: context about you. Your identities, your story bank, your second brain. That's the work we did earlier in this series. Once you've built that, your prompt can say "my audience is this, the scenario is this, the pain point is this, and by the way, pick up the George B. Thomas identity and use a story from the story bank." We covered how to build that identity layer back in Session 2 on your digital identity.
Task: The Action Plus The Why
This is where a lot of us stop too early. We give the AI an action and skip the intent behind it. The task isn't just "write this blog article." The task is the action plus the why.
Are you trying to educate? Persuade? Build trust? Motivate? "Educate someone who's skeptical" and "persuade someone who's ready to buy" send the same blog article in two completely different directions. So always pair the action with intent. "Write this to build trust with someone who's been burned by AI before." That little why steers the whole rest of the task.
Format: Tell It The Shape
Tell your AI the shape you want. Length, structure, sections, paragraph rules. "600 words, three short sections, no paragraph longer than three sentences." Headers or no headers. Emojis or no emojis.
If you don't name the shape, you get whatever shape it decides to give you. Spend five seconds on format and you save yourself ten to fifteen minutes of reformatting on the back end.
Constraints: Goal Posts And Guardrails
Constraints are the rules, and this one is huge. George had a buddy he used to work with, Remington Bagby, who always talked about goal posts and guardrails. You need both.
Most of us skip this entirely. We forget to tell the AI what not to do, and the negatives are gold. "Don't use jargon." "Don't use acronyms." "Don't be salesy." "Don't start with certain words." Those negatives cut off the generic paths you never wanted in the first place. So don't just say what you want. Say what you never want.
Examples: Show It What Right Looks Like
Show the AI what right looks like. Paste in a paragraph you wrote. Drop in a historical article you loved. Say "match this voice" or "match this style." This is the fastest path to a one-shot result, because now the AI has an example plus the direction to read it, and the output sounds like you instead of like a robot.
George doesn't paste samples much anymore, because he has identities his AI can pull from: him in training mode, him writing an article, him coaching. Same idea, just systematized.
Stack All Six And Feel The Difference
Here's what happens when you put all six together into one email-coach prompt:
Role: You're an email marketing coach. Context: my audience is solo consultants drowning in tools. Task: educate them so they feel calm and capable, not sold to. Format: 600 words, three short tips, short paragraphs. Constraints: no jargon, not salesy, one clear action step. Examples: here's a paragraph in my voice, match it.
Read that back and feel the difference from "write me a blog post." We went from a one-line vending-machine code to a structured framework that actually shapes how the AI works.
And here's the part that turns this from a trick into a system. You build that prompt once, then you hand it to your teammates. Now everybody on the team gets the same kind of result. Bake out the prompts your team uses over and over, and suddenly you're getting consistent output across multiple humans. That's a whole different game.
What Are The Three Conversation Patterns?
The three conversation patterns are the Interview Flip, the Redirect Loop, and the Mirror. They sit on top of the six-element framework and turn the AI from a one-and-done vending machine into a teammate that improves its own work.
The Interview Flip
Tell the AI: "Before you answer, ask me five questions to get the context you need." Now the AI pulls the context out of you instead of guessing. You can end almost any prompt with "interview me for any gaps I haven't given you," and watch it surface the stuff you forgot to mention.
The Redirect Loop
You never accept version one. You prompt, you review, you redirect, you refine. The whole loop is prompt, review, redirect, refine. Send it back with specific feedback every time.
The Mirror
Tell the AI: "Now put on the editor's hat, critique what you just wrote, then rewrite it." George does this with real agents. He has an editor agent named Finch who runs the critique, and a writer agent named Quinn who takes the rewrite rules and moves forward. The AI looks at its own work and improves it.
Iterate. Never accept version one.
Write that on a sticky note and stick it on your monitor. But here's the catch, and it matters: "never accept version one" does not mean typing "make it better." That gets you nothing. Vague feedback fixes nothing.
The opening is too generic. Start with a question that names their frustration.
That gets real improvement. Specificity wins the day.
The Live Demo: From Cardboard To Publishable
To make all of this concrete, George ran the same task through three layers, live on screen. The task: "write me a blog post about email marketing." Watch the output climb.
Layer one, the bare prompt. "Write me a blog post about email marketing." No role, no context, no task intent, no constraints, nothing. The output opened with "every few years, someone declares email dead, a social algorithm shifts, a new messaging app gets hot." Not terrible. But generic. The internet talking to the internet.
Layer two, the framework prompt. George added a role (email marketing coach), context (solo consultants overwhelmed by tech and short on time), a task intent (educate them, leave them calm and capable), a constraint (600 words, three short sections), and an example (this article, in my voice). The output shifted: "Let me tell you something that should make your shoulders drop. The problem was never that you aren't tech savvy enough. The problem is you've been handed too many tools and told they all matter. They don't. Most of them are noise." Now it's starting to sound like a human. On voice. And tighter, because the 600-word constraint reined it in.
Then he ran the Redirect Loop on it: "The opening is too generic. Start with a question that names their frustration. Put on the editor's hat, find the weakest sentence, and rewrite it sharper." The result: "Ever end a workday exhausted, only to realize you spent more time wrestling your tools than serving your clients?" That's a hook. The question pulls the reader in, and the AI even handed back editor's notes on the weakest sentence.
Layer three, the second-brain prompt. This is how George actually works. He told it to rewrite the article using the main brain (all the context of who he is), pick up the George B. Thomas identity, use Quinn the writing agent and the superwriter skill, pull a relevant story from the story bank, and activate mentor-and-challenger George B. Thomas. The output:
Ever finish a workday wiped out, only to realize you spent more hours fighting your tools than helping the humans who actually pay you? Here's the thing. The problem was never that you're not tech savvy enough. I've been the overwhelmed one too. Before I ever became the HubSpot guy, I was a chameleon, bouncing from bouncer to youth pastor to designer to video editor, grabbing every new tool and role I could.
There it is. A real hook, George's actual voice, and a true story from his career woven right in. Behind the build: the George Filter applied per section (clear, SEEN, SAFE, SMART), the constraints honored (600 words, three sections, three-sentence paragraphs), and Quinn even flagged an option ("if you ever want a Sidekick-branded version with a soft CTA, I can add it without disturbing the calm-and-capable tone").
Cardboard to publishable. Same task, three layers of context. That's the whole point.
Two Questions From The Room
One attendee, Eric, asked two questions during the session that are worth folding in, because they're the questions most humans have.
Is this THE framework you use for every client? Yes. The six-element framework holds no matter what George is working on, for Sidekick Strategies or for clients. It's always running in the back of his head. The second-brain layer (identities and stories) is the next level on top of it, and that's where the output really separates from everyone else's.
What's the bare minimum context? George pushed back gently here, and it's a great reframe. Stop chasing the minimum. If you're a good boss and you hire a new human, you give them everything you can so they succeed. You don't hand them the bare minimum and hope. Treat the AI the same way. Give it the maximum useful context, then pair it with the Interview Flip ("interview me for any gaps") so it closes whatever you missed. The minimum-viable mindset just gets you lipstick on a pig.
Do You Even Read The Output Anymore?
Eric asked the trust question directly: have you gotten your main brain to the point where you don't even read the output? George's honest answer is worth sitting with, because it's where the brand-protection happens.
He doesn't skip review. He listens to it.
I put an audio player on every page I create and have it read the page back to me. My eyes scan while my ears listen, and I imagine I'm somebody else reading it for the first time.
He's scanning for factual drift, the stuff the internet or the brain hasn't been updated on. Sometimes the AI says he's been in the HubSpot space for 15 years when it's been 14. Sometimes it calls segments "lists," because HubSpot changed the name. Sometimes it mentions a feature that's been sunsetted. Those are the things he hunts for.
And when he finds drift, that's a signal to go train the brain. HubSpot ships something like 18 to 28 updates a week, which would bury anyone trying to keep a brain current by hand. So Sidekick Strategies built a system: the team feeds HubSpot updates into a tool that writes HubSpot update articles, and the brain gets trained on those articles. Seven to ten new updates a day, fed in automatically. By the time George sits down to write content, the brain already has the context.
The minute it says something dumb and I don't catch it, I'm eroding my brand. Every piece of content should be building the brand.
That's the discipline behind the speed. The goal is to get content to 95 to 98 percent ready, so George is only tweaking a few tiny things before it goes live to the humans.
Why This Compounds
Here's why all of this matters. It compounds. You apply the six-element framework. You wrap the three conversation patterns around it. You build the prompt once. And now you've built something that works Monday, works Tuesday, works always.
The reason George cares about getting to 95 to 98 percent ready isn't speed for speed's sake. It's time to first value. The faster a publish-ready piece reaches a human who needs it, the faster that human gets helped. Helping humans, building the brand, and a real editing process to get there. That's the whole equation.
Your Homework Before The Next Session
One task. Two upgrades. One comparison.
- Pick one thing you do every week. Your newsletter, your client recap, your sales-call prep, any task you run on repeat.
- Rebuild it with the six elements. Role, Context (reader and you), Task plus why, Format, Constraints (load up the negatives), Examples.
- Wrap the three conversation patterns around it. Open with the Interview Flip, run the Redirect Loop with specific feedback, finish with the Mirror.
- Compare old output to new. Put last week's result next to this week's. Feel the difference, then keep the new prompt so your future self never starts from zero again.
Frequently Asked Questions
Is this the prompt framework you use for every client?
Yes. The six-element framework (Role, Context, Task, Format, Constraints, Examples) is the baseline George runs no matter the project, for Sidekick Strategies and for clients. The second-brain layer (identities and a story bank) is the next level stacked on top, and that's where the output starts to sound unmistakably like you instead of like everyone.
What's the bare minimum context I need to give AI?
Stop chasing the minimum. Treat your AI like a new hire you want to succeed: give it the maximum useful context, then end your prompt with the Interview Flip ("interview me for any gaps") so it surfaces whatever you forgot. Hunting for the floor gets you lipstick-on-a-pig output. Giving it everything you can gets you something close to publish-ready.
Do you even read AI output anymore, or do you just publish it?
George reads (and listens to) every piece. He puts an audio player on each page and has it read back to him while he scans, looking for factual drift: a wrong year, a renamed HubSpot feature, a sunsetted tool. Then he trains the brain on what it got wrong. The goal is content that's 95 to 98 percent ready, with a human still in the loop on the last mile, because every piece either builds the brand or erodes it.
What does "iterate, never accept version one" actually mean?
It does not mean typing "make it better." Vague feedback fixes nothing. It means giving specific, surgical redirects: "the opening is too generic, start with a question that names their frustration" or "put on the editor's hat, find the weakest sentence, and rewrite it sharper." Specificity is what drives real improvement.
Your Next Move
Keep going with the free series. Session 4 layered the prompt framework on top of the four-layer setup. The free 8-part Building Your AI Content System series keeps building from here, session by session. Explore the full series and grab the next live session.
Catch up on the prior sessions. Missed the setup that this framework sits on top of? Session 3 walked through setting up your AI partner the right way (custom instructions, memory, projects, and files), and Session 1 made the case for why systems beat tools. Watch those and this framework clicks even harder.
Go build the real thing. Ready to stop watching and start building? The paid AI Content System training walks you through the whole pipeline, from identity to publish, with the second-brain layer George demoed live. Want the deeper build behind that layer first? Read Second Brain Mastery: How To Build An Obsidian AI System That Actually Remembers You, then see the paid training.
References: Anthropic's prompt engineering guide and OpenAI's prompt engineering guide both back up the "examples" element with its formal term, few-shot prompting.
