24:27Ben AI · Tutorial
This Claude Second Brain Setup Will Change How You Do Sales Forever
A folder of markdown files, kept current by scheduled Claude routines, replaces the CRM as the place sales reps actually work from.
July 8thThe argument in one line.
Claude Opus 5 and Fable 5 are trained to run entire tasks autonomously, so step-by-step scripts, hard 'never' rules, and requests to self-verify now waste tokens instead of improving the output.
Read if. Skip if.
READ IF YOU ARE…
- You use Claude Code, Claude Desktop, or the API for complex multi-step tasks and want to reduce wasted back-and-forth.
- You maintain a CLAUDE.md or skill files full of hard 'never do X' / 'always do Y' rules and haven't revisited them since older models.
- You've noticed Claude Opus 5 running longer or spending more tokens than expected and want to know why.
- You want a repeatable template for briefing an AI agent on an end-to-end task instead of re-explaining context every time.
SKIP IF…
- You only send short, single-shot prompts where a full job/why/guardrails/done brief would be overkill.
- You're looking for prompt wording tricks rather than a change in how much context and autonomy you hand the model.
The full version, fast.
Anthropic's own guides for Claude Opus 5 and Fable 5 say these models perform best given a complete task instead of a step-by-step script, so Ben builds a four-part brief around that: the job, the why, the guardrails, and what done means. The other six habits follow the same logic: interview yourself before a complex task to surface unknowns, state why a request matters and not only what it is, define a hard exit condition so the model stops instead of over-running, swap blunt 'never do X' rules for instructions paired with reasons, stop asking Claude to double-check its own work since it already self-corrects, and fix a verbose or jargon-heavy voice once in a standing instruction instead of per message.
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 – 00:37
01 · Cold open
Ben claims almost everyone is prompting Claude 5 models wrong based on Anthropic's new guides and keynotes, then introduces himself as an AI agency owner and community teacher.
00:37 – 03:00
02 · Give The Whole Job
Anthropic's Opus 5 and Fable 5 guides, plus a Boris Cherny Y Combinator clip, say these models perform best given the complete task spec upfront and left to run; Ben introduces his four-part JOB / WHY / GUARDRAILS / DONE MEANS template.
03:00 – 06:21
03 · Interview Yourself First
Drawing on Anthropic's 'Knowing your unknowns' and 'Help Claude help you' articles plus an Andrej Karpathy voice-ramble tip, Ben demos an 'Interview Me' skill that questions you before a complex task, and pitches his AI Accelerator community mid-chapter.
06:21 – 07:57
04 · Why, Not Just What
Claude Fable 5's docs recommend explaining the intent behind a request, not only the instruction, using the template: 'I'm working on [larger task] for [who it's for]. They need [what the output enables].'
07:57 – 09:17
05 · Define Done
Because Claude 5 models tend to do too much rather than too little, Ben stresses explicit exit criteria and output style in every brief, backed by a worked output example.
09:17 – 10:12
06 · Reasons, Not Rules
Anthropic's context-engineering research found hard 'never do X' rules underperform instructions paired with a reason; Ben shows a real prompt-diff (an escalation policy) rewritten from a bare prohibition into a reasoned instruction.
10:12 – 11:38
07 · Avoid Double-Checking
Claude Opus 5 and Fable 5 are trained to self-verify, so asking them to double-check or spin up a verification subagent just burns tokens; Ben also flags 'think step by step,' explicit reasoning requests, and ALL-CAPS emphasis, and demos Anthropic's Rule Rewriter skill converting a CLAUDE.md's hard rules into reasoned ones.
11:38 – 13:15
08 · Fix The Voice Once
For Opus 5's drift toward jargon and verbosity, Ben sets one standing voice instruction in Claude's general settings rather than repeating it per prompt, then closes on Anthropic's 'brilliant but new employee' framing and the golden rule of prompt clarity.
Lines worth screenshotting.
- Claude Opus 5 and Fable 5 are trained specifically on executing end-to-end tasks, so giving them the complete job upfront outperforms step-by-step instructions.
- Anthropic's four-part prompting template is: THE JOB (the outcome in one or two sentences), THE WHY (who it's for and what they need), THE GUARDRAILS (what to leave alone), and DONE MEANS (the exit criteria).
- Boris Cherny, creator and head of Claude Code, says the common mistake is giving Claude overly specific step-by-step instructions instead of describing the task, guardrails, and exit criteria and letting the model run.
- Anthropic internally uses an 'interview me' skill that asks clarifying questions about a task before the model is sent off to do complex work.
- A voice ramble into a transcription tool for a few minutes surfaces more real context than typing a short prompt, because LLMs reconstruct rambling speech into something cleaner than the original.
- Claude Fable 5's documentation frames unclear requirements as 'unknowns' and says the model's biggest bottleneck is often the user's own clarity about what they want, not the model's capability.
- Stating why a request matters, not just what it is, lets Claude make the small unspecified decisions a complex task always produces the way you would have.
- Anthropic's template for stating intent is: 'I'm working on [larger task] for [who it's for]. They need [what the output enables].'
- Claude 5 models default to doing too much rather than too little, so an explicit definition of 'done' (size, scope, output style) is what actually stops runaway token spend.
- Anthropic's research found Claude 5 models respond better to instructions paired with a reason than to blunt prohibitions like 'never do X' or 'avoid Y.'
- The same reasons-not-rules rewrite applies to CLAUDE.md and skill files, not just individual prompts, and Anthropic ships a Rule Rewriter skill specifically to automate it.
- Telling Claude Opus 5 or Fable 5 to double-check its own work or use a subagent to verify adds cost without improving quality, because these models are already trained to self-correct.
- Anthropic also recommends dropping 'think step by step,' 'explain your reasoning,' and aggressive ALL-CAPS emphasis, since they can cause the model to over-trigger.
- Claude Opus 5's tendency toward jargon and verbosity is fixed once with a standing instruction in CLAUDE.md, Claude Desktop settings, or a system prompt — not by re-correcting every message.
- Anthropic frames the whole shift as: think of Claude as a brilliant but new employee who lacks context on your norms and workflows — if a prompt would confuse a colleague, it will confuse Claude too.
Seven rules for prompting Claude 5 right
PROMPTING HABITS
Claude Opus 5 and Fable 5 are trained to run whole tasks end to end, so the real shift is giving them the full job, the why, and a clear finish line instead of step-by-step scripts.
02Give The Whole Job
- Give Claude Opus 5 and Fable 5 the complete task spec upfront (the job, the why, the guardrails, and what done looks like) instead of a numbered step-by-step script, since these models are trained to execute end-to-end.
- Use a fixed four-part template for every non-trivial prompt: THE JOB (one or two sentences of outcome), THE WHY (who it's for and what they need), THE GUARDRAILS (what to leave alone), and DONE MEANS (the exit criteria).
- Save the small, well-understood tasks for quick single-shot prompts; the full framework is overkill for anything that doesn't need 10-15 minutes of autonomous work.
03Interview Yourself First
- Before sending a complex or ambiguous task, run an 'interview me' pass: have Claude ask clarifying questions first, then turn your answers into the job/why/guardrails/done brief.
- If you don't already know your own unknowns, Claude can't fill the gaps for you — it will either follow bad instructions literally or quietly default to its own judgment.
- A voice ramble into a transcription tool for a few minutes surfaces more real context than typing a short prompt, especially when cleaned up by a second pass into a structured brief.
04Why, Not Just What
- State the intent behind a request, not just the instruction, using the fill-in template: 'I'm working on [larger task] for [who it's for]. They need [what the output enables].'
- On any task complex enough to force small in-the-moment decisions, a model that understands why makes those calls the way you would.
- The habit mirrors delegating to a person: telling someone why the work matters gets better judgment calls than telling them only what to do.
05Define Done
- Claude 5 models default to doing too much rather than too little, so an explicit exit criterion (size, section count, word count) is what actually stops the token burn.
- Definition of done has two parts — what 'finished' looks like, and the output style you want it delivered in — both belong in the same brief.
- One worked example of the target output format helps the model match structure even when the prompt text alone leaves it ambiguous.
06Reasons, Not Rules
- Anthropic's own research found Claude 5 models follow instructions paired with a reason more reliably than blunt 'never do X' prohibitions.
- Rewrite guardrails from a ban into a reasoned instruction: instead of 'never mention a claim not backed by Anthropic,' say 'make sure every point is backed by Anthropic's own team so we have proof for the claim.'
- The same rewrite applies to CLAUDE.md and skill files, not just one-off prompts — hard rules written for older models are worth revisiting under this lens.
07Avoid Double-Checking
- Don't instruct Claude Opus 5 or Fable 5 to verify its own output or spin up a subagent to double-check — these models are already trained to self-correct, so the extra step adds cost with no quality gain.
- The same de-prioritization applies to 'think step by step,' 'explain your reasoning,' and aggressive capital-letter emphasis; Anthropic flags all three as no longer helping and sometimes causing over-triggering.
- Anthropic's Rule Rewriter skill can sweep an existing CLAUDE.md or skill file and convert its hard-rule phrasing into reasoned instructions automatically.
08Fix The Voice Once
- If Opus 5 keeps drifting into jargon-heavy, overly verbose responses, fix it with one standing instruction in CLAUDE.md, Claude Desktop settings, or the system prompt — not a per-message correction.
- Anthropic's own example instruction: 'Keep responses focused, brief and concise, avoid jargon and being overly verbose,' set once in general instructions and applied to every future response.
- The unifying idea behind all seven rules: treat Claude like a brilliant but new employee who lacks your context — if a prompt would confuse a colleague with no background on the task, it will confuse Claude too.
Terms worth knowing.
- Claude Opus 5
- Anthropic's top-tier Claude 5 model, built for long, high-effort, largely autonomous tasks like multi-file refactors and complex analysis.
- Claude Fable 5
- A Claude 5 family model Anthropic positions for long-horizon, high-throughput agentic work with strong first-shot correctness on complex, well-specified problems.
- Interview Me skill
- A Claude Skill that questions the user about an ambiguous task before drafting the actual work brief, used to surface requirements the user hasn't articulated yet.
- Rule Rewriter skill
- An Anthropic-provided Claude Skill that scans a CLAUDE.md or skill file and converts hard 'never/always' rules into instructions paired with the reason behind them.
- CLAUDE.md
- A project-level instructions file that Claude Code and similar agent tools read automatically at the start of every session, defining rules and context for that project.
- Over-triggering
- When emphasis techniques like ALL CAPS or repeated warnings cause a model to overreact to an instruction well beyond what was actually intended.
- Context engineering
- Anthropic's term for deliberately shaping what information, instructions, and structure a model sees, as distinct from simply wording a prompt more cleverly.
Things they pointed at.
01:08linkBoris Cherny (Claude Code creator) — Y Combinator keynote clip
03:20linkAnthropic blog — 'Knowing your unknowns' / 'Help Claude help you'
05:01linkAndrej Karpathy — X post on voice-ramble prompting
09:30toolAnthropic Rule Rewriter skill
Lines you could clip.
01:08
“You want to describe the task. You want to describe the guardrails. You want to describe the exit criteria. And then just go with the model cook and come back in a little bit.”
08:40
“Cloud five models are long running, and that means they generally don't do too little, they do too much.”
09:50
“I'm not saying never give a point that is not backed up by Anthropic. Instead, I'm saying make sure that the points that are mentioned are backed up by Anthropic's own team.”
12:58
“Think of Claude as a brilliant but new employee who lacks context on your norms and workflows.”
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.
metaphoranalogy
I recently looked at Anthropic's new guides and keynotes on prompting, and I learned that almost everyone is prompting Cloud 5 models wrong. Because Cloud Opus 5, Fable 5, and similar models work fundamentally different than previous models, we need to change the way we prompt in 2026. So in this video, I'll show you seven rules for prompting Cloud 5 models the Anthropic team actually uses to get the most out of these models and to avoid burning tokens.
Now, if you're new here, I'm Ben. I'm a three -time founder and I now run an AI agency and one of the largest AI communities for professionals and business owners where I teach how to use AI to actually automate work. Starting with the first and probably most important one, which is giving Cloud 5 models the entire job instead of prompting it step -by -step.
Now, both Entropic guides on prompting best practices for Cloud Opus 5 and Fable 5 mentioned specifically that these models tend to perform best when given the complete task specification upfront and are left to run, including for complex tasks. Now, the reason for this is simple. These models have been trained specifically on executing end -to -end tasks.
Boris Cherny also recently mentioned this in his keynote talk at Y Combinator, including a simple framework to think through when prompting these Cloud 5 models. I think a really common mistake that I see is people are using Cloud Code, they're using Cloud, and they just give away way overly specific instructions. They're like, I want you to do this, but I want you to do it in this way, this way, this way.
You must do like one, then two, then three, then four. And for modern models, that's actually really not the way to do it. You want to go a little bit higher level.
You want to describe the task. You want to describe the guardrails. You want to describe the exit criteria.
And then just go with the model cook and come back in a little bit. And I think it'll surprise you. And again, this is just not something that would have worked six months ago, but it does work today.
So essentially we want to describe the entire job that needs to be done, the why it needs to do this job, the guardrails for doing it, and of course what done looks like. So an easy way to help you improve your prompts for these types of models is always to just go through this prompting sort of framework when putting in a prompt.
I'm a big fan of these frameworks because it just makes it easy for you to go through these best practices every time you give these models a prompt. For example, I used this prompting framework together with CloudFable 5 to help me prepare for this YouTube video. So I gave it a broad job or goal.
I gave it the why, the guardrails and what done looks like. I then let it run for 10, 15 minutes and it came back with a pre. uh pre -outlined for my youtube video which was honestly extremely helpful now of course sometimes you do have smaller tasks where you might not have to be as detailed but then using these types of models is often just overkill or you should at least use them with low effort now i've put this prompting framework in the free resources in the first link in description below together with a couple of other free resources from this video But because, of course, we're not always 100 % clear on the exact specification of an entire task beforehand, especially if it's a complex one, Entropic recommends using a specific skill to get full clarity on the task beforehand.
Which brings me to rule two, which is using the interview me skill before actually sending the model off towards an end -to -end task. Basically, of course, when you're prompting a model for an end -to -end task or a complex task, it becomes harder to consider or articulate all of the little guardrails, rules and things it has to do beforehand.
But in a recent article on how the Entropic team uses Fable 5 models, they do mention it's really important to articulate and know in detail what you want out of the model in order to get the best results. That's why Entropic internally uses an interview me skill before sending one of these powerful models off to do it. a complex task.
It's basically a skill that asks you questions about the task beforehand to get the full context and then can create a brief or even the entire prompt that you can throw into a Cloud5 model. So it's designed to... find some of the unknowns in your task and bring them above surface.
For example, here I use that skill to help me plan for a personal analytics dashboard that I want to build out for myself. And as I wasn't 100 % clear on what I actually wanted or wanted in there, I used the interview me skill. So it asked me a couple of questions to get more context around my task.
And then at the end, it gave me the brief, which is essentially the prompt that we can throw into a cloud five model to actually build the dashboard. So you can see we have the job, the why, the guardrail. and what done looks like.
Now, the interview me scale will also be available in the free resource link below, but in essence, putting more effort into the pre -planning instead of... quick prompting and iterating on outputs is a little bit of the mindset shift that you need to make with these more intelligent models in my experience, because you can better spend a bit more time on creating this initial prompt and planning it.
I've personally also seen this just results in better outputs from these types of models and will also help you reduce token burn and save you time because you don't have to endlessly iterate. Now, another effective way of doing that pre -planning besides the interview skill is what also Andrej Karpathy recommends, which is basically switching on or using a voice transcription tool and just ramble on about the task for a couple of minutes.
Now, in my case, I use Whisperflow, but Cloud also has a built -in voice tool. Because through your voice, you often just get far more context out to Cloud and it's far faster. Now I do often notice with myself, for example, that becomes very organized and messy, which is also not good for these types of models.
So I sometimes use a prompt master scale that I built myself that basically gets a ramble like that and organizes and structures it a bit better. So I can then paste it into a class five model. Again, the prompt master scale will also be in the free resources.
By the way, if you enjoy this type of content, you'll really enjoy my AI Accelerator, where I just launched three new beginner to advanced cloud courses that walk you through step -by -step how to use cloud to actually automate your work as a non -technical professional or business owner. We also have unlimited one -on -one live tech help to help you with any problems you might face and a big library of skills and plugins across all business departments that you can use or customize for yourself.
We have multiple weekly Q &As with me and my team and a community with professionals and business owners actively using AI. to automate their own work or build AI business. So if you want more information, you can check it out in the link in the description below.
And if you're someone that wants a bit more help, you can also check out my AI Operator Program, which is a 30 -day one -on -one coaching program to help you get from zero to advanced really fast. So you can also check that out in the description below. Which brings me to the third rule, prompt why it needs to do this job, not just what.
The Cloud Fable 5 guide mentioned specifically that it tends to perform better when it understands the intent behind the request, especially on complex tasks. And the reason, of course, again, is because on complex tasks, the model will always run into some decisions that it has to make that you haven't pre -specified, even when you've pre -planned this with, for example, an interview skill.
And when it understands the bigger picture and why it's doing this job, it can make those small decisions. They even give a specific template you could apply in your prompts, which again, I'll add in the free resources, which is basically I'm working on the larger task for the specific person who this is for. They need.
what the output enables and with that in mind you put in the request so again it's just a little bit of a habit you want to get into when prompting these models for example in that same prompt i used that template to explain the why so i said i'm working on a video for my youtube channel on how to prompt cloud models and how it's changed the videos for non -technical professionals and business owners that are using cloud to automate their work they need practical tips examples and frameworks on how to improve their prompting not just theory And this all might feel a little bit stupid or doing too much when you first do this, but it can really improve your outputs.
And that's why I'm such a big fan of having these sort of frameworks, these prompting frameworks, because it forces you to include these things that might not come very natural. And if you think about it, the same principle really applies to giving instructions to humans. The easy thing to do is delegating work by telling someone what to do, but the most efficient way to delegate work is also telling why they're doing it and why it's important.
So rule four, which might be one of the most important ones is defining what done looks like. Cloud five models are long running and that means they generally don't do too little, they do too much. And that's why defining the exit criteria becomes so important because you might've noticed it just keeps going on.
too long and spend far too many tokens if you don't. It's also something Boris specifically mentioned in that Y Combinator talk. So there are really two aspects to this.
First, defining what done looks like and defining the output style, which is again in that same prompting framework. And it's something you really always want to try to include inside of these prompts. For example, in my case, I set a pre -outline for this video.
That means eight to 15 practical tips on how to prompt cloud models together with a specific example for each and the source you found this tip for. Again, the source you found for this tip should be backed up by Anthropic. And I added in my second brain, you can find an example output style for a video outline.
So the model can actually get an idea of what a video outline looks like. Here, of course, you can also add in a prompt of what kind of tone of voice it needs to use. Because a common problem, of course, with Opus 5 is that it can be overly verbose.
So here you can just say something like, give the output in plain English. That's easy to understand. Now, some people have been saying even examples are not that necessary anymore, but in my experience, if you have it, it does help, especially for getting a specific output style.
So rule five is swap hard rules for reasons. Now, many people are trained in using phrases like never do or avoid this, et cetera, but based on Entropic's new research and article on context engineering and a keynote note talk that they had on prompting, they've basically found that these new, more intelligent models seem to...
respond far better to instructions combined with a reason behind them. So defining guardrails is still important to do, but avoiding telling it what not to do and always adding the reason for the instruction seems to be far more effective for these more intelligent models. So when I'm defining the guardrails for this prompt, for example, I'm not saying never give a point that is not backed up by Anthropic.
Instead, I'm saying make sure that the points that are mentioned are backed up by Anthropic's own team. So we actually have proof for the claims we're making and the reason behind. Again, It might sound like a small thing, but it's more just getting into the habit of changing the way you give these guardrails a rule.
Which brings me to rule six, avoid double checking prompts in Clod 5 models. Now, this same principle, of course, also applies to your cloud .md and your scale files, which is why Entropic even launched a rule rewriter scale that helps you update your cloud .md or your scale files to swap these hard rules according to the best practices for these newer models.
For example, here I applied it to my LinkedIn writer scale and it changed some of these hard rules here to instructions with an actual reason behind them. The rule rewriter scale I'll also add in the free resources link. Telling a model to verify its output or using a sub agent to double check before giving you the final answer has been sort of a popular prompting technique for trying to reduce mistakes and hallucinations.
But cloud models like Opus 5 and Fable 5 have been specifically trained to double check and verify and fix mistakes autonomously. Meaning that specifically instructing this just adds another unnecessary verification step to the process, which adds cost without really improving. Now, besides that, there are a few other popular prompting techniques that the Entropic team recommends avoiding with these newer models, including the think step -by -step, explain your reasoning.
or using aggressive emphasis like this with capital letters because it can cause over -triggering. Now, if you want to look at all these rules in detail, the list will also be in the free resources. And the last one applies to a problem many people have been facing specifically with Cloud Opus 5, where the model starts using a lot of jargon, becomes overly verbose, and it sort of becomes unreadable, which is basically a fixed Cloud Voice Once rule.
And Tropic essentially recommends using a simple prompt inside of, for example, your Cloud .md or your Cloud Desktop. instructions or your system prompt to fix any voice issues you might have with Opus 5. They give an example in the guide again, but of course you can customize this and specify it to the specific output style you're looking for.
In the cloud desktop, for example, you can just go in general and in instructions for cloud, give the specific output style you're looking for. In my case, keep responses focused, brief and concise, avoid jargon and being overly verbose. But you can also, of course, add this in your cloud .md or even in your system prompt if you may be using projects.
So essentially, all these rules really come down to... Changing a little bit of the mindset and the way you think about prompting some of these newer models, it becomes more and more essentially like instructing a human. You don't just want to tell a human what to do.
You want to tell them why they'll need to do it. And they'll often do their work better. And I think Entropic put it very nicely in one of their articles.
Think of Cloud as a who lacks context on your norms and workflows. And the golden rule is show your prompt to a colleague with minimal context on the task and ask them to follow it. If they be confused, Clot will be too.
So essentially we have to start prompting these models more and more like humans. Again, if you want to dive deeper into AI, definitely check out my AI accelerator in the link in the description below. And if you want to learn more about how to use Clot effectively, you can check out the video here above.
The Hook
The bait, then the rug-pull.
Ben pulled Anthropic's own internal prompting guides for Claude Opus 5 and Fable 5, cross-referenced them against a Y Combinator talk from Claude Code's creator, and reduced them to seven concrete swaps: whole jobs instead of steps, reasons instead of rules, and one voice fix instead of constant re-correction.
Named ideas worth stealing.
03:00template
The Full Job Brief (JOB / WHY / GUARDRAILS / DONE MEANS)
- THE JOB
- THE WHY
- THE GUARDRAILS
- DONE MEANS
A four-part template for handing Claude Opus 5 or Fable 5 an entire task in one prompt: the outcome in one or two sentences, who it's for and why, what to leave alone, and the exact exit criteria plus output style.
Steal forany end-to-end Claude Code or agent task brief
09:30concept
Reasons-Not-Rules Rewrite
Take any hard 'never do X' guardrail and rewrite it as an instruction plus its underlying reason — e.g. swapping 'avoid escalating unless absolutely necessary' for 'escalation costs about $8 in agent time, so route only when nothing else resolves it.'
Steal forCLAUDE.md and skill-file guardrails
How they asked for the click.
VERBAL ASK
05:30product
“Become Your Company's Chief AI Officer In 30 Days — Apply Now”
Folded directly into the middle of the Rule 2 chapter as a screen-shared landing page with testimonials, rather than held for a single end-of-video ask; a second, softer CTA for the AI Accelerator and AI Operator Program repeats in the outro.
MENTIONED ON CAMERA
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
AFFILIATECommission earned if you click.
OTHER LINKSAlso linked in the description.
Visual structure at a glance.
