12:42Jay E | RoboNuggets · Tutorial
How the Top 1% Use Claude Differently
Anthropic's Claude lead engineer breaks ambitious AI use into three levers, altitude, depth, and range, plus the habit that makes all three stick.
October 17thThe argument in one line.
A more capable model raises the ceiling on what Claude can produce, but reaching that ceiling depends entirely on the context, outcome, and grading you give it, not on the model itself.
Read if. Skip if.
READ IF YOU ARE…
- You use Claude for writing, marketing, or business tasks and start most conversations from a blank chat with no saved context.
- You've switched to a newer, smarter Claude model and expected better results without changing how you prompt it.
- You want a repeatable process for getting a finished, on-voice draft out of Claude instead of a generic first pass you have to fix by hand.
SKIP IF…
- You're already running Claude Projects with saved context, outcome-based prompts, and reusable templates. This formalizes what you're doing and won't surprise you.
- You're looking for Claude API or coding-specific techniques. This is about chat-based business and content work, not development workflows.
The full version, fast.
Claude Opus 5.5, released September 22, 2026, matches Claude Fable 5.1's performance on most work at roughly 40% lower cost than Opus 5. That raises the ceiling on what Claude can do but doesn't automatically improve results, because a more capable model still needs a real brief. The video teaches FORGE, a five-step system for closing that gap: Feed Claude context once through a Project (who you are, who it's for, examples of your best work); describe the Outcome instead of the task; let Claude Reverse-interview you before it starts; Generate multiple versions and have Claude Grade them against your own criteria; then Export the winning process as a reusable template or Skill. The conclusion: a better model doesn't close the gap between people who use AI well and people who don't, it widens it.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Create a free account →Where the time goes.
00:00 – 01:31
01 · Why the same AI gives different results
Two people open the same Claude account. One gets a generic first draft, the other gets a finished, in-voice product in the same time it took the first person to ask for a rewrite. The difference isn't a secret prompt, it's a system.
01:31 – 03:08
02 · What Opus 5.5 actually changes
Claude Opus 5.5, released September 22, 2026, performs at the level of Claude Fable 5.1 on most work while costing about 40% less than Opus 5. A more capable, cheaper model raises the ceiling, but the floor still depends on the brief you give it.
03:08 – 05:42
03 · F: Feed it context
Load a Claude Project once with who you are, who it's for, and two or three examples of your best work instead of restating context every chat. Examples beat adjectives, but curate them, since Claude copies weak examples too.
05:42 – 07:29
04 · O: Outcome, not task
Describe the outcome you want, not the task, by answering who it's for, what they should do or feel afterward, and how you'll know it worked.
07:29 – 09:17
05 · R: Reverse interview
Ask Claude to interview you before starting work, one question at a time, so it surfaces gaps in your brief you didn't know were there.
09:17 – 11:25
06 · G: Generate, then grade
Ask for three genuinely different versions, have Claude grade them against the success criteria from the outcome step, then fix and finalize the winner.
11:25 – 13:08
07 · E: Export the win
Turn a workflow that worked into reusable Project instructions or a saved Skill so the whole process runs in one pass the next time.
13:08 – 13:56
08 · FORGE in action
A fast run-through of all five steps back to back on one real email, showing the full loop takes about 90 seconds once a template exists.
13:56 – 15:31
09 · What to walk away with
A better model widens the gap between people with a system and people without one. The close: pick one weekly task and run it through FORGE tonight.
Lines worth screenshotting.
- Claude Opus 5.5 matches Claude Fable 5.1's performance on most work at roughly 40% lower cost than Opus 5.
- A more capable model doesn't automatically produce better results, it raises the ceiling, and reaching it depends on what you feed the model.
- Starting every chat from a blank slate is the biggest hidden cost in how most people use Claude.
- A Claude Project should hold three things: who you are, who it's for, and two or three examples of your actual best work.
- Examples of real past work teach Claude your tone faster and more accurately than describing your style in adjectives, but Claude copies weak examples as faithfully as strong ones.
- A task has no finish line, so Claude guesses the safest, most average version of it; an outcome gives Claude something concrete to aim for.
- Before hitting enter, answering who this is for, what they should do or feel after, and how you'll know it worked turns a vague ask into a gradeable one.
- Letting Claude interview you before it starts work surfaces gaps in your own brief that you didn't know you had.
- Adding the line 'if anything I said is vague or contradicts itself, point it out' catches self-contradictions in a brief, not just missing information.
- Asking for three genuinely different versions and grading them against your own criteria gets more out of Claude than accepting the first draft.
- A more capable model is also a better critic, so the generate-then-grade step gets more valuable as the model improves, not less.
- A workflow that isn't saved as a reusable template gets abandoned by the next session, turning it into a template is what makes it stick.
- A better model doesn't close the gap between people who use AI well and people who don't, it widens it.
A better model only helps with a system
WHAT TO LEARN
Claude Opus 5.5 raised the ceiling on what the model can produce, but only a five-step system, loading context, naming the outcome, interviewing, grading, and saving the win, actually gets you there.
02What Opus 5.5 actually changes
- Claude Opus 5.5, released September 22, 2026, performs at the level of Claude Fable 5.1 on most work while costing about 40% less than Opus 5 to run.
- A more capable, cheaper model raises the ceiling on what Claude can produce, but doesn't automatically raise the floor, that still depends on what you give it to work with.
03F: Feed it context
- Starting every conversation from a blank chat forces Claude to re-derive who you are each time; loading that context once into a Project removes the repeated cost.
- A Project should hold three things: who you are (business, role, what you sell and don't sell), who it's for (a specific audience description), and two or three examples of your actual best work.
- Showing Claude examples of real past work teaches tone faster and more accurately than describing your style in adjectives, but curate the examples, since Claude copies weak ones as readily as strong ones.
04O: Outcome, not task
- A task description like 'write an email' has no finish line, so Claude defaults to the safest, most average version of the request.
- Describing the outcome instead, who it's for, what they should do or feel afterward, and how success will be measured, gives Claude a target to aim for.
05R: Reverse interview
- Instead of trying to write the perfect prompt, ask Claude to interview you first: state the goal, then have it ask clarifying questions one at a time before starting.
- People are bad at predicting what a model needs to know, but a capable model is good at spotting gaps in a brief, flipping the direction surfaces information you didn't know you were missing.
- Adding the line 'if anything I said is vague or contradicts itself, point it out' catches self-contradictions in a brief, not just missing information.
06G: Generate, then grade
- Instead of accepting the first draft, ask for three genuinely different versions of the work, explicitly ruling out the same idea reworded three times.
- Have Claude grade each version against the success criteria from the outcome step, harshly, and name what's still weak in the winner before finalizing it.
- A more capable model is also a better critic, so the generate-then-grade loop gets more value out of newer models, not less.
07E: Export the win
- A workflow that isn't saved gets abandoned by the next session, ask Claude to turn a process that worked into reusable instructions once you have an output you love.
- Save the reusable instructions into Project instructions or as a Skill so the same context, questions, and grading criteria run in one pass the next time.
09What to walk away with
- A better model doesn't close the gap between people who use AI well and people who don't, it widens it, because the ceiling keeps rising for people with a system.
- Pick one task you do every week and run it through the five-step system once, rather than trying to apply it to everything at once.
Terms worth knowing.
- Claude Projects
- A Claude feature that stores persistent context, like business info, audience, and style examples, so every new conversation inside the project starts with that context already loaded.
- Claude Memory
- A Claude setting that lets stated preferences, like tone or format rules, carry across separate conversations without being repeated each time.
- Extended thinking
- A Claude mode that lets the model reason through a problem at length before answering, instead of responding immediately.
- Claude Skill
- A saved, reusable set of instructions in Claude that can be invoked to repeat a multi-step workflow without re-explaining it every time.
- FORGE
- The five-step prompting system taught in this video: Feed it context, Outcome not task, Reverse interview, Generate then grade, Export the win.
Things they pointed at.
01:31productClaude Opus 5.5
00:40product10,000 Ultimate AI Prompts / AI Magic Prompts Vol. 2
03:08toolClaude Projects
05:20toolClaude Memory
11:55toolClaude Skills
Lines you could clip.
00:14
“One of them gets a paragraph that sounds like a LinkedIn post written by a hostage.”
02:18
“That is like your gym announcing their cutting membership fees and also somehow the treadmills now work.”
02:50
“The genius didn't fail. The brief did.”
08:55
“It's like hiring a consultant who charges nothing, has infinite patience, and doesn't sigh when you give a bad answer.”
10:45
“I have been outworked by my own laziness.”
14:45
“A better model doesn't close the gap between people who use AI well and people who don't. It widens it.”
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
metaphoranalogystory
Here's something that bugged me for months. Two people open the exact same AI. Same model, same price, same blinking cursor.
One of them gets a paragraph that . The other one gets a finished product, researched, structured, in their voice, ready to ship.
And they got it in about the time it took the first person to type, make it better, for the fourth time. Same tool, completely different results. And it's not because the second person knows some secret magic prompt.
I promise you there is no magic prompt. I've looked. I've bought two of those 10 ,000 prompts PDFs.
One of them was 9000 prompts and 1000 blank pages. The difference is that the second person isn't prompting at all. They're running a system.
And with Anthropic dropping Cloud Opus 5 .5 this month, the gap between those two people just got a lot wider. Because the smarter the model gets, the more it rewards people who know how to actually use it, and the more it quietly wastes on people who don't.
So today, I'm giving you the system. I call it Forge. Five steps.
Every single one of them you can use tonight, on the free plan or paid. And by the end of this video, you'll never type, write me a blog post about X again. Well, you might, but you'll feel bad about it.
Quick context so we're on the same page. On September 22nd, Anthropic released Cloud Opus 5 .5. It's the first model in their new 5 .5 family.
And there are two things about it that matter for you. Number one, Anthropic says it performs at the level of Cloud Fable 5 .1, their top -tier model. on most work.
Fable is the one that was built for the heavy stuff. Serious coding, serious knowledge work. Number two, it costs about 40 % less to run than Opus 5 on typical workloads.
So let me translate that into normal human language. You're getting top shelf performance at a lower price. That is like .
Now, here's the part nobody's saying out loud. A more capable model does not automatically give you better results. It gives you a higher ceiling.
Whether you reach that ceiling depends entirely on what you give it to work with. Think of it like hiring. If you hire a genius and on day one you say, Do marketing stuff.
You're going to get generic marketing stuff. Very well -written generic marketing stuff. But generic.
The genius didn't fail. The brief did. That's what 99 % of people are doing with clothes right now.
They hired a genius and handed them a sticky note. Forge is how you write the brief. Let's go.
F is feed it context. Here's the uncomfortable truth. Most people start every single chat from scratch.
New chat, blank slate, hi Claude, I run a small business. Every. Single.
Time. That's like going to the same doctor for 5 years and every appointment she goes, and you are? The 1 % do the opposite.
They load context once. And then every conversation starts already knowing who they are. The easiest way to do that is projects.
You create a project and inside it you put three things. One, who you are. Your business, your role, what you sell, what you don't do.
Two, who it's for. Your audience. And I mean really describe them.
Not small business owners. More like solo founders, mostly 28 to 45, overwhelmed, allergic to jargon, already tried three productivity apps and quit all of them. Three, and this is the one nobody does, what good looks like.
Upload two or three examples of your best past work, your best email, your best script, and your best client proposal. Because here's the thing, you can spend an hour describing your writing style in adjectives, punchy but warm, professional but not stiff, and Claude will nod politely and give you something that sounds like a bank.
Or you can just show it three things you actually wrote, and it gets it instantly. Examples beat adjectives. Every time.
Quick story. When I first set this up, I uploaded my best scripts as examples. And then I read them again.
And I realized two of them were, let's say, not my best. So Claude started matching the energy of a girl who clearly filmed at 1am. So lesson learned.
Curate your examples. Claude will copy your worst habits with the same enthusiasm as the best ones. And if you have memory turned on, Claude can also carry preferences across conversations.
So things like, I hate bullet point heavy answers, or always give me the short version first, stick without you repeating yourself. Worth two minutes to set up. Genuinely, one of the highest return two minutes you'll spend this year.
O is outcome, not task. Most prompts describe a task. Write an email, summarize this report, give me ideas for a video.
The problem is that a task has no finish line. Cloud has no idea what done looks like, so it guesses. and its guess is the safest, most average version of the thing.
The 1 % describe the outcome instead. What happens after this thing exists? Watch the difference.
Task version, write an email to my list about my new course. Outcome version. I need an email that gets people who've been on my list for over six months and have never bought anything to click through to the course page.
They've seen a lot of launches from other people and they're skeptical. Success means they feel like I'm talking to them specifically and the email doesn't feel like a launch at all. Under 200 words.
No exclamation marks. I'm serious about the exclamation marks. Same request, but now Claude knows who's reading, what they're feeling, what counts as a win, and what the constraints are.
The second email is going to be dramatically better. And not because Claude got smarter between the two prompts. Because you gave it a target.
Here's the simple version you can steal. Before you hit enter, answer three questions. Who is this for?
What should they do or feel after? How will I know it worked? If you can't answer those, honestly, Claude can't either.
And that's not an AI problem. That was already a problem. The AI just made it visible.
R is reverse interview. And if you only take one thing from this entire video, take this one. Instead of trying to write the perfect prompt, You let Claude interview you first.
It looks like this. I want to mention your goal. Before you do anything, ask me the questions you'd need answered to do this really well.
Ask them one at a time and do not start the actual work until I say go. That's it. That's the hack.
And I know it sounds too simple, but think about what's actually happening. You're not good at predicting what Claude needs to know. Nobody is.
You don't know what you don't know. But Claude is extremely good at spotting the gaps in a brief. So you flip the direction.
You stop guessing and let it pull the information out of you. And the questions it asks are often better than the ones you'd think to answer. The first time I tried this, it asked me, what's the one thing your audience believes that's stopping them from buying?
And I sat there for a full minute because I realized I had never actually thought about that for three years. It's like , which honestly is more than I can say for some consultants. Pro tip, for anything important, add this line at the end.
If anything I said is vague or contradicts itself, point it out. Because sometimes the problem isn't that you left information out, it's that you told it two opposite things and it didn't notice. G is generate, then grade.
Here's how most people work with AI. Ask for one thing, get one thing, and then either accept it or start the no -make -it -more -like, you know, spiral.
The 1 % never ask for one thing. They ask for options and then they make Claude judge them. Step one, give me three genuinely different versions of this.
Not the same idea, reworded three times. Three different angles. That last line matters.
If you don't say it, you'll get the same email in three slightly different outfits. Step two, and this is where it gets fun. Now grade each version against the success criteria I gave you.
Be harsh. Tell me which one wins and what's still weak about it. This works because of what we set up in step O.
Remember the success criteria? Now Claude has something real to measure against. It's not grading on vibes.
It's grading on your definition of done. Step three, take the winner, fix the weaknesses you just named, and give me the final version. Three steps, maybe two extra minutes.
And the difference in quality is honestly kind of embarrassing because it means for years I was accepting first drafts from a machine. that will write a second draft instantly and for free and never once complained about it. I have been outworked by my own laziness.
And with Opus 5 .5 specifically, this step pays off even more. A more capable model is also a better critic. It's better at spotting what's weak, which means the grade and fix loop gets you further per round than it used to.
If you want to go one step further, Turn on extended thinking for the grading step. Let it actually reason through the critique instead of rushing into an answer.
It's the difference between a teacher skimming your essay and one who actually reads it. And E, the last one is export the win. This is the step that turns the 1 % into the 0 .1%.
Here's the thing about everything we just did. It's great, but if you have to do it from scratch every time, you'll stop doing it by Thursday. I know you will, because I did.
So whenever a workflow actually works, when you get an output you love, you ask Claude one more question. That worked really well. Turn everything we just did, the context, the questions you asked, the criteria, the grading into reusable instructions I can use next time.
And now you have a template. You paste that into your project instructions or you save it as a skill if you're using those. And next time, the whole forge process runs in one go.
You're not re -explaining yourself. You're just running your system. That is the part that actually changes your week.
Because every time you do this, you're not just getting one good output. You're building a library of workflows that do exactly what you need in your voice to your standard. After a month, you have 10 of them.
After three months, you basically have a small team. A small team that doesn't need coffee, doesn't need slack, and has never once replied all to the whole company. That's the whole idea behind this channel, by the way.
One person running like a company. Not because you work harder, but because the stuff you figure out once keeps working for you after you close the laptop. Alright, let's run the whole thing once, fast, so you can see it as one flow.
F. I'm inside my project, it already knows my business, my audience, and it's got my three best emails as examples. O.
I describe the outcome, not the task. Who it's for, what they should do, what done looks like. R.
I tell it to interview me first. It asks four questions, one of them I genuinely hadn't thought about. Again, I'm starting to think that's a point.
G. Three versions, graded against my criteria, winner fixed. E.
And I save the whole process as a reusable template, so next time this takes about 90 seconds. So, forge. Feed it context, outcome not task, reverse interview, generate, then grade, export the win.
Here's what I want you to actually walk away with. The model is going to keep getting better. Opus 5 .5 is better than what came before it.
And whatever comes next will be better than this. But a better model doesn't close the gap between people who use AI well and people who don't. It widens it.
Because the ceiling keeps rising and only the people with a system are climbing toward it. You don't need to be technical to be on the right side of that gap. You just need to stop treating clothe like a research bar and start treating it like the smartest new hire you've ever had.
If you want to try it, pick one thing you do every week. Just one. And run it through Forge tonight.
Then come back and tell me in the comments what the reverse interview asked you. I read them and I'm genuinely curious what questions it's pulling out of people. And if you want the next step, subscribe.
Because in the next video, I might show you how to take those saved workflows and chain them together so one request kicks off a whole process. I'll see you there.
The Hook
The bait, then the rug-pull.
Two people can open the exact same Claude account and get wildly different results, and it has nothing to do with a secret prompt. This breakdown covers the FORGE system, and why a smarter, cheaper Claude Opus 5.5 makes the gap between the two even wider.
Named ideas worth stealing.
01:15acronym
FORGE
- Feed it context
- Outcome, not task
- Reverse interview
- Generate, then grade
- Export the win
A five-step system for prompting Claude that replaces one-shot task requests with project-loaded context, an explicitly defined outcome, an interview step before work starts, a generate-and-grade loop, and a reusable exported template.
Steal forany repeated Claude workflow: marketing copy, client proposals, research briefs, launch emails
How they asked for the click.
VERBAL ASK
14:55subscribe
“Just one. And run it through Forge tonight... And if you want the next step, subscribe.”
Soft CTA tied to a concrete action, pick one weekly task, plus a comment prompt asking what the reverse interview asked, then a teaser for a follow-up video on chaining saved workflows.
FROM THE DESCRIPTION
OTHER LINKSAlso linked in the description.
Visual structure at a glance.
