Big Idea

The argument in one line.

Anthropic's newest Claude models need far fewer explicit rules than before, so Anthropic deleted 80% of Claude Code's system prompt and shipped a built-in /doctor skill that applies the same trimming logic to a user's own CLAUDE.md, skills, and settings.

Who This Is For

Read if. Skip if.

READ IF YOU ARE…

You use Claude Code regularly and haven't revisited your CLAUDE.md file or skill list in months.
You've been stacking rules and constraints into CLAUDE.md or skills because that's what older Claude models needed.
You want a free, built-in way to audit what's bloating your Claude Code context instead of guessing.

SKIP IF…

You've never used Claude Code or don't maintain a CLAUDE.md file — there's nothing here to act on yet.
You're looking for new Claude Code features or a roadmap, not a cleanup workflow for what you already have.

TL;DR

The full version, fast.

An Anthropic engineer on the Claude Code team wrote that the newest Claude models got capable enough that Anthropic deleted over 80% of Claude Code's system prompt with no measurable drop in coding evaluations, because heavily-constrained rule lists and redundant CLAUDE.md instructions were mostly compensating for older, less capable models. Anthropic packaged the fix into a built-in /doctor skill: a read-only audit that finds unused skills and MCP servers, trims a bloated CLAUDE.md, and pushes always-loaded guidance into lazy-loaded sub-files instead. The core shift is replacing rigid rule lists with a stated goal plus the model's judgment, and loading context only when it's actually needed. Run live, /doctor found 113 personal skills eating roughly 10,000 tokens of context every session just to be listed, a quarter of them never used.

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 – 00:35

01 · Anthropic Engineer's New Rules

Cold open: the creator flags a post from Tariq, an Anthropic engineer on the Claude Code technical team, promising it changes how you should use Claude Code — including a built-in skill almost nobody talks about.

00:35 – 01:23

02 · What Context Engineering Really Means

Your typed prompt is only a small slice of what Claude actually receives — system prompt, skills, CLAUDE.md (global + project), and saved memory all get assembled together, and how well you engineer that mix determines the results.

01:23 – 02:07

03 · Why 80% Got Deleted

Anthropic's engineering team saw a large jump in how well the newest Claude models (Opus 5, Sonnet 5, Fable 5) follow instructions, so they removed over 80% of Claude Code's system prompt with no measurable loss in coding evaluations.

02:07 – 02:41

04 · The Hidden /doctor Skill

Anthropic packaged its findings into a built-in Claude Code skill called /doctor that right-sizes a user's skills and CLAUDE.md file; it's read-only first, reports back, then asks permission before changing anything.

02:41 – 03:04

05 · How The Health Check Works

The health check covers nine areas, starting with install health, duplicate or leftover installs, broken settings files, and broken or colliding agent definitions.

03:04 – 03:21

06 · How to Find Dead Weight

/doctor checks real usage counters and session transcripts to find skills, MCP servers, or plugins that are installed but never actually used, then flags them for removal.

03:21 – 04:05

07 · How to Trim CLAUDE.md

/doctor checks CLAUDE.md for repetition and flags information Claude could figure out on its own (directory layout, tech stack, build commands) instead of being hard-coded — since the file loads, and costs tokens, every session.

04:05 – 04:36

08 · How to Use Lazy Loading

Move always-loaded guidance out of CLAUDE.md into a subdirectory or skill; treat CLAUDE.md like a table of contents or index pointing to other files, not a dumping ground.

04:36 – 05:16

09 · How to Fix Permission Settings

/doctor flags scripts that block on every tool call, shows what's already eating your context budget before you type, confirms your Claude Code version, and checks that auto mode is your default.

05:16 – 06:02

10 · Why Over-Constraining Hurts Results

Reviewing real transcripts, Anthropic found contradictory rules stacked into CLAUDE.md and skills; those guardrails once prevented worst-case failures, but newer models handle ambiguity fine using surrounding context and judgment instead.

06:02 – 06:41

11 · How Skills Replace Memory Bloat

Claude Code used to rely on CLAUDE.md as its main source of memory; artifacts and skills now give it more ways to load and share context across sessions without bloating the main file.

06:41 – 07:18

12 · How to Give Claude Judgment

The old pattern was long numbered rule lists; the new pattern is stating the goal and letting Claude use its own judgment to figure out the best way to get there.

07:18 – 07:57

13 · How to Use Progressive Disclosure

Load the right context only when needed: Anthropic moved code review and verification into on-demand skills, and Claude Code can use tool search to find the right tool only when it's called for.

07:57 – 08:15

14 · How to Structure File Trees

Structure CLAUDE.md and skill files like a tree — a top-level file that tells Claude where to look, and sub-files for specific tasks — so only the necessary information loads, and nothing more.

08:15 – 09:40

15 · My Real /doctor Results

Live on his own machine: install is healthy, but /doctor finds 113 personal skills eating roughly 10,000 tokens of context every session just to list, 25 never used, before ending on a CTA to his Skool community.

Atomic Insights

Lines worth screenshotting.

Anthropic deleted more than 80% of Claude Code's system prompt for its newest models and saw no measurable drop in coding evaluations.
The built-in /doctor skill runs a full, read-only health check on a Claude Code setup and asks permission before changing anything.
A CLAUDE.md file loads on every single session, so every extra line in it is a token cost paid over and over, not once.
Contradictory rules stacked into CLAUDE.md and skills, like 'always add documentation' next to 'never add documentation,' force the model to do extra reconciliation work.
Newer Claude models handle ambiguity well enough that many explicit guardrails can be deleted and replaced with surrounding context and judgment, with no loss in quality.
Skills and artifacts now give Claude Code memory options beyond CLAUDE.md, so information no longer has to be crammed into one file to persist across sessions.
Progressive disclosure means loading context only at the moment it's needed — Anthropic moved code review and verification out of the system prompt and into on-demand skills.
Tool search lets Claude Code look up which tool fits a job only when needed, instead of loading every tool definition into context up front.
One creator's own /doctor run found 113 personal skills eating roughly 10,000 tokens of context every session just to be listed, a quarter of them never used.
The best CLAUDE.md file reads like a table of contents or index pointing to other files, not a repository containing every rule and fact.
Giving Claude a goal and trusting its judgment now produces the same results as a long numbered list of dos and don'ts, faster and cheaper.
As of this video, the /doctor skill worked in the Claude Code terminal but was not available in the desktop app.
Repeating an instruction across multiple files no longer improves compliance the way it did with older Claude models — it just adds redundant tokens.

Takeaway

Your CLAUDE.md file is costing you tokens every single session

CONTEXT CLEANUP

Newer Claude models need far fewer explicit rules than before, so trimming CLAUDE.md, deleting unused skills, and giving Claude a goal instead of a rulebook now produces the same results faster and cheaper.

02What Context Engineering Really Means

Context engineering means deliberately shaping everything Claude receives beyond your prompt: system prompt, skills, CLAUDE.md files, and saved memory all get assembled together.
How well you engineer that mix, not just what you type, determines how optimized and successful your results are.

03Why 80% Got Deleted

Anthropic's engineering team noticed a large jump in how well the newest Claude models follow instructions.
They removed over 80% of Claude Code's system prompt for these newer models with no measurable loss in coding evaluations.
That means these models need far less explicit hand-holding than earlier generations to produce the same quality of output.

04The Hidden /doctor Skill

Anthropic packaged its findings into a built-in Claude Code skill called /doctor that right-sizes a user's skills and CLAUDE.md file.
/doctor is read-only first: it scans your setup, reports what it found, and asks permission before making any changes.
As of this video, /doctor worked in the Claude Code terminal but was not available in the desktop app.

05How The Health Check Works

The health check covers nine areas, starting with install health, duplicate or leftover installs, and broken settings files.
It also flags broken or colliding agent definitions, since two agents claiming the same responsibility can produce inconsistent behavior.

06How to Find Dead Weight

/doctor checks real usage counters and session transcripts to find skills, MCP servers, or plugins you never actually use.
Anything identified as dead weight gets flagged for removal so it stops bloating your context on every session.

07How to Trim CLAUDE.md

/doctor checks your CLAUDE.md for repetition or redundancy with other files on your system.
It flags information Claude could figure out on its own, like directory layout or build commands, instead of having it hard-coded.
CLAUDE.md loads on every single session, so the slimmer that file is, the less you pay in tokens and time on every run.

08How to Use Lazy Loading

Lazy loading means moving always-loaded guidance out of CLAUDE.md and into a subdirectory or skill that only loads when relevant.
Treat CLAUDE.md like a table of contents or index pointing to your other files, instead of cramming everything into one place.

09How to Fix Permission Settings

/doctor looks for scripts that block on every tool call and reports what's already eating your context budget before you type.
It confirms you're on the latest Claude Code version and checks that auto mode, not constant permission prompts, is your default.

10Why Over-Constraining Hurts Results

Anthropic's team found contradictory rules stacked into CLAUDE.md and skills, like 'always add documentation' next to 'don't add documentation.'
Conflicting instructions force the model to do extra reconciliation work instead of just acting on your intent.
Guardrails were once needed to avoid worst-case failures with older models, but newer models can use surrounding context and judgment instead, with the same results, faster and cheaper.

11How Skills Replace Memory Bloat

Claude Code used to rely on CLAUDE.md as its main source of memory, information, and guidance.
Artifacts and skills now give Claude additional ways to load and share context across sessions without bloating the main file.

12How to Give Claude Judgment

The old pattern was long rule lists: do steps one through five, don't do these ten things.
The new pattern is to state the goal and let Claude use its own judgment to figure out the best way to get there.

13How to Use Progressive Disclosure

Progressive disclosure means loading the right context at the right time instead of stuffing everything into the prompt upfront.
Anthropic moved code review and verification out of the always-on system prompt and into their own on-demand skills.
Tool search lets Claude look up short tool definitions and call the right one only when needed, instead of loading every tool definition into every session.

14How to Structure File Trees

Structure your CLAUDE.md and skill files like a tree: a top-level file that tells Claude where to look, and sub-files for specific tasks.
The goal across every one of these fixes is the same: load only the necessary information, and nothing more.

15My Real /doctor Results

One real /doctor run found 113 personal skills eating roughly 10,000 tokens of context every session just to be listed, already over budget before any work starts.
25 of those 113 skills had never once been used, and the report named the specific extensions burning the most tokens.
/doctor also checks lazy loading, your Claude Code version, and auto-mode settings, then asks how you want to proceed before changing anything.

Glossary

Terms worth knowing.

Context engineering: Deliberately shaping everything Claude receives beyond your typed prompt — system prompt, CLAUDE.md, skills, and saved memory — to get the most reliable output.
CLAUDE.md: A project- or account-level file Claude Code reads at the start of every session for standing instructions, directory layout, and conventions.
/doctor skill: A built-in Claude Code skill that audits a user's setup (skills, CLAUDE.md, settings, permissions) and reports cleanup suggestions before making any changes.
System prompt: The baseline instructions Anthropic ships with Claude Code itself, separate from anything a user writes into CLAUDE.md.
Skill: A packaged set of instructions or a process Claude can call up on demand, keeping specialized context out of the main CLAUDE.md file.
Lazy loading: Keeping guidance out of the always-loaded CLAUDE.md and instead putting it in a subdirectory or skill that only loads when it's relevant.
Progressive disclosure: Surfacing information or tools to Claude only at the moment they're needed rather than upfront in every context window.
Tool search: A mechanism where Claude looks up which tool fits a task from short definitions, rather than every tool's full definition being loaded by default.
MCP server: A connected service that gives Claude access to extra tools or data sources, which can also bloat context if left installed and unused.
Dead weight: Skills, MCP servers, or plugins that are installed but rarely or never used, identified by usage counters and session history.

Resources

Things they pointed at.

00:04linkTariq's Claude Code context engineering article (X/Twitter) ↗

Quotables

Lines you could clip.

01:40

“We removed over 80% of Claude Code's system prompt for all of the newest models with no measurable loss in our coding evaluations.”

stat-driven, source-backed hook line→ TikTok hook↗ Tweet quote

02:07

“They've actually already put all of the best practices into a skill called the Claude doctor skill. This ships with Claude.”

reveals a free, hidden built-in feature→ IG reel cold open↗ Tweet quote

06:41

“Before you would give Claude all these rules... but now all you have to do is give Claude judgment.”

clean before/after framing of the core thesis→ newsletter pull-quote↗ Tweet quote

08:42

“I have 113 personal skills, so they eat about 10,000 tokens of context in every single session just to list them. That's already over budget.”

concrete, personal, slightly self-deprecating number→ TikTok hook↗ 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.

analogy

I just stumbled across this post from one of the engineers at Anthropic, and this is Tariq, and he's on the technical team of Claude Code. And if you've been using Claude Code for any amount of time, this is worth paying attention to. It's going to make your systems faster.

It's going to make things cost less, and it's gonna make everything more efficient. It's kind of completely changed the way I think about using Claude code. I'm even going to show you a skill that ships with Claude code that nobody is talking about, and I'm honestly not sure why.

So Tariq wrote this article titled The New Rules of Context Engineering for Quad five Models, and I know that sounds too technical, and maybe it is. And basically, this all just has to do with how we train Quad, and it seems like right now, most people are just doing it wrong. But it's really not your fault.

Like, the models have changed and have gotten so much better over time that if you haven't thought about these things for a while, they're really worth revisiting. So the way we prompt Claude has fundamentally changed. So when you send a message to Claude, the prompt is only a small part of the context it gets.

Much of your context is assembled from your system prompt, any of the skills you've created, your Claude MD files both at the global level and the project level, any memory that Claude has saved, and any other sources that you give it. So when I use the word context engineering, that's essentially what it means.

How do you engineer the context that you are giving Claude to be the most optimized, most efficient, and most successful? And so the way that you do this makes a big impact on the results that you generate when you're using quad code or building your own agents. So this is the part that I thought was really interesting.

Tariq says that most recently the engineering team at Anthropic noticed a large jump in the way we prompt the newest generation of quad models. So again, like Opus five, Sonnet five, Fable five, all the fives. They said that they removed over 80% of quad code system prompt for all of the newest models with no measurable loss in our coding evaluations.

That basically means that these models have just gotten so insanely smart that you don't need to give them nearly as much information as you had to before in order to get a really impactful output. So the rest of this article just basically goes to break out what they've learned about prompting this brand new class of models and how you can utilize it to update your own context engineering.

What's pretty cool is they've actually already put all of the best practices into a skill called the Claude doctor skill. This ships with Claude. You can use this inside of the terminal.

I checked to see it in the desktop app. It didn't look like it was available there, but you can use the command slash doctor in Claude code to right size your skills and your Claude MD file. So I wanted to just give you this at a high level so you could basically understand what the slash doctor skill actually even does.

And so I just asked Claude. I just said, hey, like, tell me simply at a high level what this skill does. So the slash doctor skill is a full health check of your quad code setup.

It's read only first, so it's going to scan everything. It reports back to you, and then it asks you before actually making any changes. So that's key because you don't want quad just doing a whole bunch of your stuff on your system that you don't understand.

And so there are nine core things that it looks at. So the install health, any duplicate or leftover installs, any broken settings files, or any broken or colliding agent definitions. Any deadweight, this one is huge.

Like, are there any skills or MCP servers or plug ins that you never actually use? It's actually going to look at your own usage counters and your session transcripts. It's just going to remove things that you're not actually using so you don't bloat all of the context.

These next two are huge, and they have to do with cleaning up your quad MD file. So it's basically going to check to make sure there's no repetition or redundancy anywhere else on your system. It's going to look to see if it can just trim your core quad MD file.

It's going to look to see if you put too much context into your QuadMD file when really like the session could just figure it out itself. Like how your directories are laid out, what your tech stacks are, are there any standard build commands. Like your QuadMD file loads every single session and so you pay for it every time.

And so the slimmer this file could be, the better. And there's a couple of ways that you can architect your files that are going to be significantly more efficient and save you time and save you on tokens. And number five is really the way to do that.

This idea of lazy loading. So always loaded guidance that would be better as a subdirectory or a skill.

And so now instead of cramming as much as you possibly can into this CloudMD file, the right way to do it is keep this file, like, really high level. Almost think of it as, like, chapters or a glossary or an index that references all of your other files. And so it basically just says, like, hey.

Here are the skills that you have access to. Here's what each one does. Like, you now know what to do.

It's going to look to see if you have any scripts that block your loop every tool call. It's going to actually look at your context weight to see what's eating your context budget before you type anything. So that's super helpful.

It's just gonna make sure you're on the latest version of Claude. You all probably are because this thing updates automatically and any permission fixes.

So it's gonna make sure that auto mode is your default so that Quad isn't constantly bothering you for feedback. So I'm just gonna go ahead and run it just to see the type of output it gets because, again, it's going to create this consolidated clean this up question and, like, a little report for me so I can see where I'm, like, bloating my system, and then I can decide what I want Claude to change and what I wanna keep.

So while that's cooking, let me just go back to this article because I think it's really interesting. And so one of the core things that they found is that at this point, like, the Anthropic engineering team was over constraining Claude code both through the system prompt and in our Claude MD files and skills. So there were a couple of examples that when they would, like, read through their transcripts of how they use Claude, they would actually see these conflicting messages that say, like, hey.

Like, leave all the documentation in or, hey. Like, don't add any documentation. And so it was saying that the system is pretty smart and, like, Cloud can actually generally interpret the user's prompts and what they wanna do, but it still has to do a little bit of extra work when it has, like, two competing instructions.

And so, again, if you've been using Cloud for a while, this is worth looking at because it seems like these are just kind of like remnants from a past way of coding with these LLMs. Like, they were once needed to avoid these worst case scenarios, like, so giving Claude a whole bunch of these guardrails and constraints, but they have found that since they can actually delete many of them and let the model use any of the surrounding context and its own judgment, and they're literally getting the same results, and it's way faster and way cheaper.

Additionally, Quad Code now has many more tools than it originally had. So Quad used to rely on QuadMD as a source of memory, information, and guidance.

But now you have that memory, but you also have artifacts and you have skills. So Quad can use these to create new ways of loading and sharing context across sessions. So if you've been creating skills for a while, you understand it's like a skill is a set set of rules or instructions or a process that you want to follow.

And so instead of coding everything into your MD files, you just say, hey, Claude, when you need to like search YouTube, like go use this skill. And then all of the context is loaded in that skill, and it doesn't bloat the rest. And so again, it's this idea of what we were doing before and what we are doing now.

And so there's just a couple of the core ones that I think are the most important. Like, before you would give Claude, like, all these rules. Like, do, like, one, two, three, four, five, and, like, don't do these 10 things.

Right? But now all you have to do is give Claude judgment. Let it know what you're trying to accomplish and say, hey.

You figure out the best way to do this. So one of the other core takeaways from this was this idea of using progressive disclosure instead of putting all of your information upfront. Before, because Claude was focused on coding, the system prompt included detailed information on how to do code review and verification.

These were not always needed, but when they were, it was crucial information, so they kept it despite it not always being needed. But since then, Claude Code has gotten way more confident at using progressive disclosure, which just means loading the right context at the right time.

So calling the information that you need when you need it, and not calling it when you don't. So for example, they moved the verification and code review into their own skills that Claude code could selectively call again only when needed. You can also do this for tools.

So instead of like loading all of your tools at once, just use tool search with some basic definitions on what your tools do. So Quad will search for those definitions, figure out what tool is right for the job, and call on it only then. I think I touched on this before, but like that same way of thinking can be applied to your quad and skill files.

And so a common myth is that you want to make these a central repository for every known practice that you might run into. But instead consider having a tree of files that can be loaded at the right time. So basically the top level QuadMD or skill file that tells Quad where to search for things and then all of the sub files for any specific individual task.

So it's only loading the necessary information and nothing more. And so really this all boils down to just simplify everything. And so thanks, Tarik, for that.

Let's go back through and just see what the doctor skill found on my own machine just so I can give you an idea of the types of things that you might find. I did run this the other day, so we should be in like a pretty reasonable place, but I still think it's interesting nonetheless. So the install is healthy.

I'm on the latest version. Nothing is broken. The real finding is SkillBlow.

Shocking. I have a 113 personal skills, so they eat about 10,000 tokens of context in every single session just to list them.

So that's already over budget. I should probably cut those down. 25 of those skills have never once been used.

So it goes ahead and lists all the different types of skills and plugins that I never used. The bulk of my token usage does look like it's coming from these 88 personal skills. Okay.

That's great. It looked at the 50 most recent sessions. Went ahead and listed any extensions that I'm not using.

I already ran this dedupe, and I trimmed the CloudMD file, so that's actually looking pretty good. But you can see here that checks two and three basically look at your overall CloudMD or your system prompt to figure out can we make it like, more streamlined and more optimized. It checks your lazy loading, checks to make sure which version you're on.

It checks auto mode, and gives you any warnings, and then basically asks how you wanna proceed. If you wanna learn how to use Claude code and get access to all my resources, just check the link in the description. If you wanna see a skill that takes all of the guesswork out of prompting Claude, check out this video right here.

I'll see you over there.

The Hook

The bait, then the rug-pull.

An engineer on Anthropic's own Claude Code team just published the new rules of context engineering — and admitted the team quietly deleted over 80% of Claude Code's system prompt with zero measurable loss in coding performance. Buried in the same article is a built-in skill, /doctor, that runs the identical cleanup on your own setup.

Frameworks

Named ideas worth stealing.

02:41list

The /doctor nine-point health check

Install health
Duplicate/leftover installs
Broken settings files
Broken/colliding agent definitions
Dead weight (unused skills/MCP/plugins)
CLAUDE.md redundancy
Lazy-loading opportunities
Context budget/version check
Permission/auto-mode settings

The nine areas the built-in /doctor skill scans before proposing any cleanup to a Claude Code setup.

Steal forauditing any Claude Code power-user setup before it silently bloats context

06:41concept

Rules to Judgment

Replace long enumerated do/don't lists in CLAUDE.md and skills with a stated goal, trusting the model's judgment to get there.

Steal forwriting leaner CLAUDE.md files and skill instructions

CTA Breakdown

How they asked for the click.

VERBAL ASK

09:28product

“If you wanna learn how to use Claude code and get access to all my resources, just check the link in the description.”

Soft spoken CTA at the very end pointing to his own paid Skool community (description link) plus a pointer to a related video — no hard sell mid-content.

MENTIONED ON CAMERA

00:04linkTariq's Claude Code context engineering article (X/Twitter) ↗

FROM THE DESCRIPTION

PRIMARY CTAWhere the creator wants you to go next.