Big Idea

The argument in one line.

Anthropic cut over 80% of Claude Code's system prompt with no measurable performance loss, proving that the smallest context that gets the job done beats stacking up instructions a newer model no longer needs.

Who This Is For

Read if. Skip if.

READ IF YOU ARE…

You run Claude Code with a CLAUDE.md file, custom skills, or project instructions that have accumulated over months.
You want a repeatable way to audit and trim an AI setup instead of guessing what's safe to delete.
You're building multi-step AI workflows and want evidence that instructions are actually reaching the steps that need them.

SKIP IF…

You've never set up a CLAUDE.md or skills and have no existing instructions to audit yet.
You're looking for a first-prompt tutorial rather than a cleanup process for an existing setup.

TL;DR

The full version, fast.

Anthropic removed over 80% of Claude Code's system prompt for newer models like Opus 5 and Fable 5 and measured no drop in coding performance, because instructions written for weaker models become restrictions once a model improves. The fix is minimal context: keep only what a model can't infer or look up, state the intent behind a request instead of a rigid procedure, and load detailed instructions only when the task actually calls for them. Claude Code's /doctor command audits an existing setup for unused skills, duplicate instructions, and stray config files pulling in extra tokens. Running that audit after every model upgrade keeps a setup lean instead of letting it accumulate rules it no longer needs.

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 →

Chapters

Where the time goes.

00:00 – 03:30

01 · Anthropic's New Context Rules

Anthropic removed over 80% of Claude Code's system prompt for newer models with no measured performance loss, and why rules built for older models become restrictions as models improve.

03:30 – 04:13

02 · Hyperagent (Sponsor)

Sponsor segment for Hyperagent, an agency workspace tool for managing client AI agents.

04:13 – 06:01

03 · Testing Claude With Fewer Instructions

A side-by-side test of a heavily-instructed setup versus almost no instructions, plus research on why longer context can hurt performance even when the information is retrievable.

06:01 – 07:04

04 · What Context To Keep

What's worth keeping in permanent instructions versus what a model can look up itself, and why stating intent beats dictating procedure.

07:04 – 09:43

05 · Auditing My Setup With /doctor

Running Claude Code's /doctor command to audit skills, memory, and duplicate instructions, and what the resulting report flagged.

09:43 – 10:51

06 · Fixing My YouTube Workflow

Using Claude to audit a real YouTube production repo and surface the smallest change that would speed up the idea-to-script pipeline.

10:51 – 12:11

07 · Scheduling AIOS Audits

Turning the audit into a recurring habit, including putting it on an automatic schedule.

12:11 – 12:22

08 · Final Thoughts

Closing thoughts on AI operating systems, human consistency, and team-wide syncing.

Atomic Insights

Lines worth screenshotting.

Anthropic deleted more than 80% of Claude Code's system prompt for models like Opus 5 and Fable 5 and reported no measurable loss on their coding evaluations.
Rules written to compensate for an older, weaker model become unnecessary restrictions once the model improves, so a rule that helped Opus 4 can actively hold back a newer model.
The best context is the smallest amount that produces the highest quality result at the cheapest cost, and minimal does not mean short if a detail is genuinely needed.
A 2025 study of five models across math, QA, and coding tasks found performance dropped between 13.9% and 85% as input length grew, even when the needed information stayed retrievable.
Loading MCP tool definitions only when needed cut token usage 85% in one Anthropic test and raised MCP evaluation scores from 49% to 74% on Opus 4 and 79.5% to 88.1% on Opus 4.5.
A version of a task run with no CLAUDE.md and no skills produced better-organized, more professional output than the fully branded, heavily-instructed version of the same task.
Explaining the intent behind a request, rather than dictating a rigid step-by-step procedure, gives a model room to make good judgment calls on the parts you didn't specify.
Claude Code's /doctor command audits skills, memory, plugins, settings, and installation, listing each skill's estimated token footprint and how many times it's actually been used.
An audit can catch instructions loaded from config files outside the active project, meaning a setup can be paying token cost for rules the user didn't know were still active.
An audit of 25 video folders with a finished script found only 3 had a dedicated sources file, showing that research evidence usually wasn't reaching the writing step in a reusable form.
Treat an instruction-and-skill audit as routine maintenance: monthly if a workflow changes often, quarterly if it's stable, and always after adopting a new model generation.

Takeaway

Less instruction, better Claude Code results

CONTEXT AUDIT

Anthropic's own data shows that trimming instructions, not adding them, is what improved Claude Code's results, and a regular /doctor audit is how to find what's safe to cut.

01Anthropic's New Context Rules

Anthropic removed over 80% of Claude Code's system prompt for newer models like Opus 5 and Fable 5 and measured no drop in coding performance, which means instructions tuned for older, weaker models can become dead weight once a model improves.
A rule that compensates for a model's past weakness, like a hard limit on comment length, outlives its usefulness once the model gets smarter, so it's worth re-testing old constraints instead of assuming they still help.
Instructions tend to accumulate the same way bugs do: a model makes a mistake, you add a rule, and that rule survives multiple model upgrades even after the underlying problem has been solved.

03Testing Claude With Fewer Instructions

Running the same task with a full custom instruction set versus almost no instructions can reveal that the stripped-down version produces better-organized output, because heavy guardrails can crowd out a model's own judgment.
The recommended standard is the smallest set of high-signal information that gets the result you want, and minimal doesn't mean short if a detail genuinely matters to the task.
A 2025 study found performance on math, QA, and coding tasks dropped 13.9% to 85% as input length grew, even when the needed information was technically still retrievable, so padding context has a real cost beyond token price.
Loading tool definitions only when needed instead of upfront cut token usage 85% in one test and raised MCP evaluation accuracy from 49% to 74% on Opus 4 and 79.5% to 88.1% on Opus 4.5.

04What Context To Keep

The information worth keeping in permanent instructions is what a model can't infer or look up on its own: your voice, your goals, your audience, and how your business actually works.
Explaining the intent behind a request, not just the steps, helps a model make good judgment calls on open-ended parts of a task, the way an experienced designer still needs to know who a project is for without a rigid checklist.
If the same instruction is duplicated across multiple files or skills, it's worth checking whether every copy is still needed rather than assuming redundancy is harmless.
A detailed procedure that only applies to one specific task is better loaded only when that task comes up, not carried in every session regardless of what's being asked.

05Auditing My Setup With /doctor

Claude Code's /doctor command audits skills, memory, plugins, settings, and installation, listing each skill's estimated token footprint and how many times it's actually been used.
An audit can catch config files loaded from unrelated locations outside the current project, meaning a setup can be paying token cost for instructions the user didn't realize were still active.
Duplicate or repeated sections across multiple instruction files are concrete, measurable waste an audit can quantify in both characters and estimated tokens.
After flagging problems, the audit can offer to fix them directly, including correcting skills with naming or formatting errors that were silently preventing their descriptions from being read.

06Fixing My YouTube Workflow

Asking an AI assistant to audit its own working repo and name the smallest change that would speed up a process, without editing anything yet, surfaces concrete handoff gaps rather than vague advice.
A verification rule added for one real reason can become redundant busywork once an earlier step in the same pipeline already found and recorded the same evidence in a reusable file.
Counting how often a specific artifact actually exists across past output is a fast way to measure whether a process is being followed consistently or just in theory.

07Scheduling AIOS Audits

Treat an instruction-and-skill audit as a recurring task, roughly monthly if the workflow changes often or quarterly if it's stable, rather than a one-time cleanup.
Every time a new model version is adopted, it's worth re-running the audit, because each model interprets the same settings, skills, and instruction files slightly differently.
Once an audit routine is trusted, it can be put on a schedule so recommendations arrive automatically instead of needing to be triggered manually each time.

Glossary

Terms worth knowing.

Context engineering: Deciding what information (system prompt, skills, CLAUDE.md files, memory) a model sees for a given task, as distinct from writing a single one-off prompt.
System prompt: The baseline rulebook that ships with Claude Code and sits underneath a user's own instructions and loaded context on every request.
CLAUDE.md: A project- or user-level instructions file that Claude Code reads automatically, used to state what a project is for and flag gotchas to watch for.
/doctor: A Claude Code slash command that audits a local setup, reporting skill token usage, duplicate instructions, broken files, and settings.
MCP (Model Context Protocol): A standard for giving Claude Code access to external tools, whose definitions can be loaded only when a task needs them instead of upfront.

Resources

Things they pointed at.

00:19linkThe New Rules of Context Engineering for Claude 5 Generation Models (Anthropic blog post)

03:30toolHyperagent ↗

07:04tool/doctor (Claude Code slash command)

Quotables

Lines you could clip.

02:13

“The best context is the smallest amount of context that gets you the highest quality result at the cheapest cost.”

concise standalone thesis for the whole video→ TikTok hook↗ Tweet quote

05:39

“It's kind of like giving an experienced designer the checklist that you would give to a kid making their first presentation.”

vivid analogy that explains the argument in one line→ IG reel cold open↗ Tweet quote

01:05

“Every six months, you should delete all of your skills.”

provocative one-liner attributed to Claude Code's creator→ newsletter pull-quote↗ Tweet quote

The Script

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

So Anthropic Engineers just 10x 'ed Claude code by deleting over 80 % of its system prompt. But they also changed everything that we know about prompting. Because a lot of us were told that the best way to get Claude to behave the way we want it to was to give it as much information and context as possible.

But with this change, almost all of your information is actually slowing Claude down and slowing down its thinking speed. So... Today I'm going to show you how to fix your projects using the engineering approach from Anthropic, so let's get into it.

Alright, so Thoric, one of the engineers at Anthropic, recently published an article called The New Rules of Context Engineering for Cloud 5 Generation Models. His team removed over 80 % of Cloud Code's system prompts for models like Opus 5 and Fable 5 and 5 .1 and future models, and they reported no measurable loss on their evaluations of these models.

The system prompt is basically the rulebook that comes with Cloud Code, so your own instructions and whatever context gets loaded for the task sit right alongside of the system. impromptu. So they're saying that they could cut most of that rulebook and keep the performance that they had measured.

And in my own setup, I used their guidance to review the instructions that I had accumulated and clean up how my skills work together. And this goes right alongside with Boris Cherny, the creator of CloudCode said himself, which was basically like every six months, you should delete all of your skills. And we've been told that skills are what makes CloudCode much smarter, and much more consistent.

Now Anthropic describes that rules they needed for older models became too restrictive as Claude actually just got better and better. One example was comments in code, because they used to put a hard limit on how much Claude could write, even though some projects needed more explanation. Their replacement now tells it to follow the style of the surrounding code.

I think it's pretty easy to end up with the same problem in your own setup. You know, Claude makes a mistake, you add a rule, and then a few months later, you've got that rule in your Claude .md file or in a skill, maybe a slightly different version of it somewhere else, and you don't actually need that rule because you keep upgrading the model, but the instructions still stay in there, and the model might not need all of that extra context now.

I recently did a video about prompting Fable 5 .1. I talked about the idea of defining to the model what done looks like.

You basically need to give it the outcome that you're after, why you're doing it, and the constraints that it needs to respect. And then you just have to get out of its way, leave it some room to actually do the work. So for this video, I asked Claude to look at how I get from a YouTube idea to a script that I could film or an outline that I could film about, and I wanted it to find a small improvement and point me to the file so that I could check the recommendation for myself.

So I'll show you guys that demo of me actually doing this later in the video. Now, Anthropic's Applied AI team recommends using the smallest amount of that gets the result you want.

And I turn that into sort of my own saying, which is the best context is the smallest amount of context that gets you the highest quality result at the cheapest cost. They also explicitly say that minimal doesn't necessarily mean short. If the model needs a detail to do the job, then obviously you need to keep that detail in there.

And there's research looking at what happens when you add more inputs. A paper published in 2025 tested five models on math, question answering, and coding, and performance dropped as the input got longer, even when the models could retrieve the relevant input. information.

The reported drops ranged from 13 .9 % to 85%, depending on the model and the task. Now, those were obviously different models, and that was over a year and a half ago, and these were controlled experiments, so those percentages don't tell us what will happen when you trim your own Cloud .md, for example. But Anthropic also tested loading MCP tool definitions only when needed.

They reported an 85 % reduction in token usage in their example, and their MCP evaluation scores improved from 49 % to 74 % for Opus 4, and from 79 .5 % to 88 .1 % for Opus 4 .5. So there are measured results behind this exact idea.

And in that case, Cloud still had access to the tools. It just didn't have to load all of those instructions up front. Now in your own setup, you can use a slash doctor.

And when I recently did this, the doctor report gave me this workflow audit with specific instructions and handoffs that I need to clean up. So I'm also going to show you guys that in just a sec. Real quick guys, a message from today's sponsor, HyperAgent.

If you're building agents for clients, most of the work lands around the agent because each client's data has to stay separate, their staff needs to use it without breaking it, and somebody has to be obviously paying for those runs. So what HyperAgent does is they give every client their own workspace. So the agents, the skills, the memories, and the documents all live inside one perimeter.

So nothing from one client shows up in the context of another agent. And your builders work in there with full edit access, and the client's own staff can get a role that runs the agents but can't touch the system prompts or swap the models or accidentally delete a skill or anything like that. So if that sounds useful for your agency, then check out HyperAgent through the link in the description.

Huge thanks to HyperAgent for sponsoring this part of the video. Now let's get back to it. Now I have tried some pretty extreme versions of this just experimenting with my own AI operating system.

You know, I've duplicated it before. I removed the Cloud .md. I've removed my skills and I've tried off, you know, a bunch of different prompts to see them do the same job.

So one example that I recently did was I took this interview with Boris Cherny and I wanted to turn it into a student resource guide. Now my full AIOS setup that I was used to running every day, it made the prettier document because it had my branding, it had a proper heading, it had links that I normally include, and it had just like more specific context about me, all that kind of stuff.

And the fresh version looks... little bit messier it wasn't as pretty but its structure was much better it had better details it organized the interview into ideas and it added time stamps and it just felt more professional and it felt like a better guide to actually give to a student so what i had was a polished guide from one version and a content structure that i preferred from the other version now obviously i still want my branding so what i would do is i would build a new skill out of this and i would say hey you know you should be using this as the header and this as the font and this as the footer and add these logos and things like that but i'm going to leave it up to you to actually structure the content in the way that you see fit because because I've gave you the, basically the intent of turn this into a student resource guide.

Because in the pretty version that had my skill already, I had so many guardrails and so many instructions that I was basically just like confining the model and I wasn't letting it truly be creative and give me the output that it thought was best because it was reading my skill and it had to follow my instructions. And those instructions were helpful back when we had like Opus 4, but now that we have Fable 5 .1, it's able to just do a better job on its own.

It's kind of like giving an experienced designer the checklist that you would give to a kid making their first presentation. the designer still needs to know who it's for and what you're trying to accomplish. But if you dictate the placement of every box and every sentence before they've even seen the material, you might prevent them from coming up with something that they might add that you would actually like more.

And that's how I think about reviewing these instructions. I want to keep the details that help Claude understand the job and what makes it specific to me, and then test it whether it still needs the procedure that I've written around it. So when I talk about the context part of my four C's of an AI OS, this is the kind of information I mean.

I mean my voice, I mean my goals, I mean my audience, and the way - that my business works. Claude can't just infer all of that or look it up online.

So that's what I have to supplement to it. Now, it's also important to explain why you're asking for something, the intent behind it. So if I'm making a resource guide for students, that helps Claude make decisions about using, you know, analogies and simple explanations and using examples.

And I'd keep that context when I clean up the skill. But a long folder listing is something it can probably look up. So if the same instruction is repeated in multiple different places, I'd check whether I need all of those copies or if I could just clean them.

up. And if a detailed procedure only applies to when I'm editing a video, I'd rather load it in when I'm actually editing a video. And this is all about getting out of the model's way, which still means you have to be clear about what it's allowed to do.

So I'd keep the Cloud .md useful as a starting point, obviously telling Cloud things like what the project is for, giving it any of the gotchas and the landmines to watch out for, and point it toward the deeper instructions. So now my whole video editing process doesn't need to be there when I'm asking for help with a script.

So inside Cloud Code, you can run slash doctor to look through a lot of the stuff happening. locally that you probably don't spend much time thinking about. Settings, config files, tedious stuff like that.

So I ran slash doctor in a new session and it checks my skills, my memory, my plugins, my settings, and my installation. And you can see in this table right here, it lists each skill where it's installed and how many recorded uses it has since installation and an estimate of the tokens that it's listing takes up. That basically just means like the YAML front matter.

It uses that YAML front matter to find the skill, which is different from loading in the full instruction when it actually decides to run it and read the whole thing. For example, agent builder is estimated at 111 tokens. and proveit is 103 tokens, and both show zero recorded uses here.

The report flagged 18 older skills with no recorded use, but kept the ones that I'd only created a few days ago, so it's taking their age into account as well. It also checked 50 recent sessions from this project and says that it also looked at lifetime counters across 252 startups and skill and plugin dispatches across 223 transcripts.

So I'd still review that list before I decided to turn anything off or clean anything up or delete anything, especially if I do use a skill through a different tool, for example, but these the usage records available to Cloud Code in this audit. And then it also checks for things like duplication.

So in mine, the project map was repeating things already covered by the routing map. There were also repeated sections for templates, references, and archiving files. And it identified about 2 ,250 characters of cuts with an estimated saving of 563 tokens.

And you can see it breaks out different places that memory is actually coming from. You know, I've got my project level Cloud .md, my user -wide instructions, the auto memory index, and my Cloud local MD for instructions specific to this checkout on my machine. The local file was just an empty template estimated at about 35 tokens.

Now that's obviously not a ton, but the point I'm trying to make here is that it will look through a lot. of the stuff that you probably don't even think about ever. It also found the cloud .md from the separate video project in my desktop folder that loads into Herc 2, which is roughly another 2 ,000 tokens just by its estimate here.

So even if you've cleaned up the file inside your project, there can be instructions coming from somewhere else that you don't even know you're technically paying for. So the setup checks caught two broken skills as well. One had the wrong file name, the other had a formatting problem in its front matter, the little description block at the top like we talked about.

So cloud wasn't actually properly seeing its description. And those files have been corrected because because at the end of the doctor audit, it says, hey, do you want me to just fix all this?

It also looks at things like your installation. It looks at things like your automatic updates and it checked the settings behind auto mode permissions and my hooks. And real quick, guys, if you're learning how to build with AI, I've got a free community with a bunch of resources and skills and courses on building AI systems and agents.

And you can also ask questions and work through the stuff with other people who are also behind you or in front of you. So if you want to join, the link is in the description. Let's get back to the video.

All right. So after running slash doctor, I asked Claude Code to actually look at my YouTube workflow. And this is the prompt that I used.

I said, audit this repo and show me the smallest change that would make the path from a raw YouTube idea to a filming ready script faster. Use the systems already here, cite the exact files you would reuse and do not edit anything yet. You can see that it mapped the steps from finding an idea through research to an outline, to a script and the teleprompter file that I can actually plug in and read while filming.

Then it looked at how the work gets passed from one step to the next. The main recommendations about research, my outline skill could gather facts, but it didn't record. saving the evidence in a file that ScriptSkill knew how to check.

And my ScriptSkill had a hard rule telling it to verify the facts again. Now obviously I put that there because I wanted verification. I wanted to make sure that it wasn't making up facts or information.

But if we've already found the primary source and checked what it says, I'd like the next step to have that work in front of it. You can see the audit counted 25 video folders with a script, but only three with a file specifically called sources .md. Now this doesn't mean that all the other videos had no research, but it does show that this particular way of handing over the evidence was inconsistent.

So it suggested three edits across the two existing skills. And I was able to once again, just say, yep, that sounds good. Let's go ahead and make those changes.

So now that you understand how you can use Claw to audit your own setup, I would make this something that you're doing regularly, maybe once a month, you know, if you're adding a bunch of skills and you're changing your workflow all the time, or maybe once a quarter, if the setup stays pretty stable. And honestly, what I would recommend is every time a new model drops and you switch into it, you should probably just run this sort of audit anyways, because every model is going to interpret your settings and your skills and your instruction files just a little bit differently.

What's really cool once you've done this a few times and you start to have more trust in the doctor or in like your audit prompts, you can start putting these checks on a schedule. So for this kind of local setup check, you could use a local schedule task that can see the same files and settings inside of the cloud desktop app.

And you could just leave it open. And then, you know, once a month, it'll run that doctor and it will run these different audits. And then you can wake up to recommendations that you can basically just approve or.

deny. We've talked a lot about AI OSs or AI second brains and a big pain point with those is on the humans, right? Human consistency is the issue.

Sometimes we forget to update things or we, you know, put duplicate entries in there or we forget to delete things, stuff like that. So if you can actually have your AI OS work with you on that cadence of automating this stuff and cleaning this stuff, then you're creating this really powerful sort of like self -improving, self -cleaning AI operating system, which is obviously something that there's a lot of value in because once you have your own personal AI OS figured out, then you can start to figure out how do you have your team stay synced up and how you have everyone working together rather than everyone duplicating work and building their own AIOSs.

So anyways, that is going to do it for this one. I hope that you guys enjoyed or you learned something new. And if you did, please give it a like.

It helps me out a ton. And as always, I appreciate you guys making it to the end of the video and I'll see you on the next one. Thanks everyone.

The Hook

The bait, then the rug-pull.

Anthropic quietly rewrote the rulebook for how to prompt its newest models, cutting most of Claude Code's own system prompt with no drop in performance. This breaks down what that means for anyone's CLAUDE.md and skills, and runs the same kind of audit against a real production setup.

Frameworks

Named ideas worth stealing.

02:13concept

The Best-Context Rule

The best context is the smallest amount of information that gets the highest quality result at the cheapest cost. Minimal doesn't mean short if a detail is actually needed for the task.

Steal fordeciding what stays in any CLAUDE.md or skill file

CTA Breakdown