Anthropic Engineers Just 10x'd Everyone's Claude Code — Transcript
Full transcript
- 0:00So, Anthropic engineers just 10x Claude
- 0:01code by deleting over 80% of its system
- 0:04prompt, but they also changed everything
- 0:06that we know about prompting. Because a
- 0:07lot of us were told the best way to get
- 0:09Claude to behave the way we want it to
- 0:11was to give it as much information and
- 0:12context as possible. But with this
- 0:14change, almost all of your information
- 0:16is actually slowing Claude down and
- 0:17slowing down its thinking speed. So,
- 0:19today I'm going to show you how to fix
- 0:21your projects using the engineering
- 0:22approach from Anthropic. So, let's get
- 0:24into it. All right, so Thoric, one of
- 0:26the engineers at Anthropic, recently
- 0:27published an article called The New
- 0:28Rules of Context Engineering for Claude
- 0:305 generation models. His team removed
- 0:32over 80% of Claude code system prompt
- 0:34for models like Opus 5 and Fable 5 and
- 0:365.1 and future models, and they reported
- 0:38no measurable loss on their evaluations
- 0:40of these models. The system prompt is
- 0:42basically the rulebook that comes with
- 0:43Claude code. So, your own instructions
- 0:45and whatever context gets loaded for the
- 0:46task sit right alongside of the system
- 0:49prompt. So, they're saying that they
- 0:50could cut most of that rulebook and keep
- 0:52the performance that they had measured.
- 0:53And in my own setup, I used their
- 0:55guidance to review the instructions that
- 0:56I'd accumulated and clean up how my
- 0:58skills work together. And this goes
- 0:59right alongside with Boris Cherny, the
- 1:01creator of Claude code said himself,
- 1:02which was basically like every 6 months
- 1:04you should delete all of your skills.
- 1:06And we've been told that skills are what
- 1:07makes Claude code much smarter and much
- 1:09more consistent. Now, Anthropic
- 1:10describes that rules they needed for
- 1:11older models became too restrictive as
- 1:13Claude actually just got better and
- 1:15better. One example was comments in code
- 1:17because they used to put a hard limit on
- 1:18how much Claude could write even though
- 1:20some projects needed more explanation.
- 1:21[music] The replacement now tells it to
- 1:22follow the style of the surrounding
- 1:24code. I think it's pretty easy to end up
- 1:25with the same problem in your own setup.
- 1:27You know, Claude makes a mistake, you
- 1:29add a rule, and then a few months later
- 1:30you've got that rule in your Claude.md
- 1:32file or in a skill, maybe a slightly
- 1:34different version of it somewhere else,
- 1:35and you don't actually need that rule
- 1:37because you keep upgrading the model,
- 1:38but the instructions still stay in
- 1:40there. And the model might not need all
- 1:41of that extra context now. I recently
- 1:43did a video about prompting Fable 5.1. I
- 1:45talked about the idea of defining to the
- 1:47model what done looks like. You
- 1:49basically need to give it the outcome
- 1:50that you're after, why you're doing it,
- 1:52and the constraints that it needs to
- 1:53respect. And then you just have to get
- 1:54out of its way. Leave it some room to
- 1:56actually do [music] the work. So, for
- 1:57this video, I asked Claude to look at
- 1:58how I get from a YouTube idea to a
- 2:00script that I could film [music] or an
- 2:01outline that I could film about, and I
- 2:02wanted it to find a small improvement
- 2:04and point me to the file so that I could
- 2:05check the recommendation for myself. So
- 2:07I'll show you guys that demo of me
- 2:08actually doing this later in the video.
- 2:10Now, Anthropic's applied AI team
- 2:11recommends using the smallest amount of
- 2:13context that gets the result you want,
- 2:15and I turn that into sort of my own
- 2:16saying, which is the best context is the
- 2:19smallest amount of context that gets you
- 2:20the highest quality result at the
- 2:22cheapest cost. They also explicitly say
- 2:24that minimal doesn't necessarily mean
- 2:26short. If the model needs a detail to do
- 2:28the job, then obviously you need to keep
- 2:29that detail in there. And there's
- 2:30research looking at what happens when
- 2:32you add more input. A paper published in
- 2:342025 tested five models on math,
- 2:36question answering, and coding, and
- 2:38performance dropped as the input got
- 2:39longer, even when the models could
- 2:41retrieve the relevant information. The
- 2:43reported drops ranged from 13.9% to 85%
- 2:46depending on the model and the task.
- 2:48Now, those were obviously different
- 2:49models, and that was, you know, over a
- 2:50year and a half ago, and these were
- 2:52controlled experiments, so those
- 2:53percentages don't tell us what will
- 2:54happen when you trim your own Claude 3
- 2:56MD, for example. But Anthropic also
- 2:57tested loading MCP tool definitions only
- 2:59when needed. They reported an 85%
- 3:02reduction in token usage in their
- 3:03example, and their MCP evaluation scores
- 3:05improved from 49% to 74% for Opus 4, and
- 3:09from 79.5 to 88.1% for Opus 4.5. So
- 3:13there are measured results behind this
- 3:14exact idea. And in that case, Claude
- 3:16still had access to the tools. It just
- 3:18didn't have to load all of those
- 3:20instructions up front. Now, in your own
- 3:21setup, you can use a slash doctor, and
- 3:23when I recently did this, the doctor
- 3:24report gave me this workflow audit with
- 3:26specific instructions and handoffs that
- 3:28I need to clean up. So I'm also going to
- 3:29show you guys that in just a sec. Real
- 3:30quick, guys, a message from today's
- 3:31sponsor, HyperAgent. If you're building
- 3:33agents for clients, most of the work
- 3:35lands around the agent. Because each
- 3:37client's data has to stay separate,
- 3:39their staff needs to use it without
- 3:40breaking it, and somebody has to be
- 3:42obviously paying for those runs. So what
- 3:44HyperAgent does is they give every
- 3:46client their own workspace. So the
- 3:47agents, the skills, the memories, and
- 3:49the documents all live inside one
- 3:51perimeter, so nothing from one client
- 3:53shows up in the, you know, the context
- 3:55of another agent. And your builders work
- 3:56in there with full edit access, and the
- 3:58client's own staff can get a role that
- 4:00runs the agents, but can't touch the
- 4:01system prompts or swap the models or
- 4:03accidentally delete a skill or anything
- 4:05like that. So, if that sounds useful for
- 4:07your agency, then check out Hyper Agent
- 4:08through the link in the description.
- 4:10Huge thanks to Hyper Agent for
- 4:11sponsoring this part of the video. Now,
- 4:12let's get back to it. Now, I have tried
- 4:13some pretty extreme versions of this,
- 4:15just experimenting with my own AI
- 4:16operating system. You know, I've
- 4:17duplicated it before, I've removed the
- 4:19Claude 3 MD I've removed my skills, and
- 4:20I've tried off, you know, a bunch of
- 4:22different prompts to see them do the
- 4:23same job. So, one example that I
- 4:24recently did was I took this interview
- 4:26with Boris Cherney, and I wanted to turn
- 4:28it into a student resource guide. Now,
- 4:29my full AIOS setup that I was used to
- 4:31running every day, it made the prettier
- 4:33document because it had my branding, it
- 4:35had a proper heading, it had links that
- 4:36I normally include, and it had just like
- 4:38more specific context about me, all that
- 4:39kind of stuff. And the fresh version
- 4:41looked a little bit messier, it wasn't
- 4:42as pretty, but its structure was much
- 4:44better. It had better details, it
- 4:46organized the interview into ideas, and
- 4:47it added timestamps, and it just felt
- 4:49more professional, and it felt like a
- 4:50better guide to actually give to a
- 4:52student. So, what I had was a polished
- 4:53guide from one version and a content
- 4:55structure that I preferred from the
- 4:57other version. Now, obviously, I still
- 4:58want my branding, so what I would do is
- 5:00I would build a new skill out of this,
- 5:01and I would say, "Hey, you know, you
- 5:02should be using this as the header and
- 5:04this as the font and this as the footer
- 5:05and add these logos and things like
- 5:06that. But, I'm going to leave it up to
- 5:08you to actually structure the content in
- 5:09the way that you see fit because I've
- 5:11gave you the basically the intent of
- 5:13turn this into a student resource
- 5:14guide." Because in the pretty version
- 5:16that had my skill already, I had so many
- 5:18guardrails and so many instructions that
- 5:20I was basically just like confining the
- 5:21model, and I wasn't letting it truly be
- 5:23creative and give me the output that it
- 5:24thought was best because it was reading
- 5:26my skill, and it had to follow my
- 5:27instructions. And those instructions
- 5:29were helpful back when we had like Opus
- 5:314, but now that we have Fable 5.1, it's
- 5:33able to just do a better job on its own.
- 5:35It's kind of like giving an experienced
- 5:36designer the checklist that you would
- 5:37give to a kid making their first
- 5:39presentation because the designer still
- 5:40needs to know who it's for and what
- 5:41you're trying to accomplish. But, if you
- 5:43dictate the placement of every box and
- 5:45every sentence before they've even seen
- 5:47the material, you might prevent them
- 5:48from coming up with something that they
- 5:50might add that you would actually like
- 5:51more. And that's how I think about
- 5:52reviewing these instructions. I want to
- 5:54keep the details that help Claude
- 5:55understand the job and what makes it
- 5:57specific to me, and then test it whether
- 5:59it still needs the procedure that I've
- 6:00written around it. So, when I talk about
- 6:01the context part of my four C's of an AI
- 6:04OS, this is the kind of information I
- 6:06mean. I mean my voice, I mean my goals,
- 6:08I mean my audience, and the way that my
- 6:09business works. Claude can't just infer
- 6:11all of that or look it up online, so
- 6:13that's what I have to supplement to it.
- 6:15Now, it's also important to explain why
- 6:17you're asking for something, the intent
- 6:18behind it. So, if I'm making a resource
- 6:20guide for students, that helps Claude
- 6:22make decisions about using, you know,
- 6:24analogies and simple explanations and
- 6:25using examples. And I'd keep that
- 6:27context when I clean up the skill. But,
- 6:29a long folder listing is something it
- 6:30can probably look up. So, if the same
- 6:32instruction is repeated in multiple
- 6:34different places, I'd check whether I
- 6:36need all of those copies or if I could
- 6:37just clean them up. And if a detailed
- 6:39procedure only applies to when I'm
- 6:41editing a video, I'd rather load it in
- 6:43when I'm actually editing a video. And
- 6:44this is all about getting out of the
- 6:46model's way, which still means you have
- 6:47to be clear about what it's allowed to
- 6:49do. So, I'd keep the Claude on MD useful
- 6:51as a starting point. Obviously, telling
- 6:52Claude things like what the project is
- 6:53for, giving it any of the gotchas and
- 6:55the landmines to watch out for, and
- 6:57point it toward the deeper instructions.
- 6:59So, now my whole video editing process
- 7:01doesn't need to be there when I'm asking
- 7:02for help with a script. So, inside
- 7:04Claude code, you can run {slash} doctor
- 7:06to look through a lot of the stuff
- 7:07happening locally that you probably
- 7:08don't spend much time thinking about,
- 7:10settings, config files, tedious stuff
- 7:12like that. So, I ran {slash} doctor in a
- 7:14new session, and it checked my skills,
- 7:16my memory, my plugins, my settings,
- 7:18[music] and my installation. And you can
- 7:19see in this table right here, it lists
- 7:21each skill, where it's installed, and
- 7:22how many recorded uses it has since
- 7:24installation, and an estimate of the
- 7:26tokens that it's listing takes up. That
- 7:28basically just means like the YAML front
- 7:30matter. It uses that YAML front matter
- 7:32to find the skill, which is different
- 7:34from loading in the full instruction
- 7:36when it actually decides to run it and
- 7:37read the whole thing. For example, Agent
- 7:39Builder is estimated at 111 tokens, and
- 7:42Prove it is 103 tokens. And both show
- 7:44zero recorded uses here. The report
- 7:46flagged 18 older skills with no recorded
- 7:48use, but kept the ones that I'd only
- 7:50created a few days ago. So, it's taking
- 7:52their age into account as well. It also
- 7:53checks 50 recent sessions from this
- 7:55project and says that it also looked at
- 7:57lifetime counters across 252 startups
- 7:59and skill and plug-in dispatches across
- 8:02223 transcripts. So, I'd still review
- 8:04that list before I decide to turn
- 8:05anything off or clean anything up or
- 8:06delete anything, especially if I do use
- 8:08a skill through a different tool for
- 8:09example. But, these are the usage
- 8:11records available to Claude code in this
- 8:13audit. And then it also checks for
- 8:14things like duplication. So, in mine,
- 8:16the project map was repeating things
- 8:18already covered by the routing map.
- 8:19There were also repeated sections for
- 8:21templates, references, and archiving
- 8:22files, and it identified about 2,250
- 8:25characters of cuts with an estimated
- 8:27saving of 563 tokens. And you can see it
- 8:29breaks out different places that memory
- 8:31is actually coming from. You know, I've
- 8:32got my project level Claude MD, my
- 8:34user-wide instructions, the auto memory
- 8:36index, and my Claude local MD for
- 8:39instructions specific to this checkout
- 8:41on my machine. The local file was just
- 8:43an empty template, estimated at about 35
- 8:45tokens. Now, that's obviously not a ton,
- 8:46but the point I'm trying to make here is
- 8:47that it will look through a lot of the
- 8:49stuff that you probably don't even think
- 8:50about ever. It also found the Claude MD
- 8:52from the separate video project in my
- 8:53desktop folder that loads into Herc too,
- 8:56which is it roughly another 2,000 tokens
- 8:58just by its estimate here. So, even if
- 8:59you've cleaned up the file inside your
- 9:01project, there can be instructions
- 9:02coming from somewhere else that you
- 9:03don't even know you're technically
- 9:05paying for. So, the setup checks caught
- 9:06two broken skills as well. One had the
- 9:08wrong file name, the other had a
- 9:10formatting problem in its front matter,
- 9:11the little description block at the top
- 9:13like we talked about. So, Claude wasn't
- 9:14actually properly seeing its
- 9:15description. And those files have been
- 9:17corrected because at the end of the
- 9:18doctor audit, it says, "Hey, do you want
- 9:20me to just fix all of this?" It also
- 9:21looks at things like your installation,
- 9:23it looks at things like your automatic
- 9:24updates, and it checked the settings
- 9:25behind auto mode, permissions, and my
- 9:27hooks. And real quick guys, if you're
- 9:29learning how to build with AI, I've got
- 9:30a free community with a bunch of
- 9:31resources and skills and courses on
- 9:33building AI systems and agents, and you
- 9:35can also ask questions and work through
- 9:36the stuff with other people who are
- 9:38also, you know, behind you or in front
- 9:39of you. So, if you want to join, the
- 9:40link is in the description. Let's get
- 9:42back to the video. All right. So, after
- 9:43running /doctor, I asked Claude Code to
- 9:45actually look at my YouTube workflow.
- 9:47And this is the prompt that I used. I
- 9:48said, "Audit this repo and show me the
- 9:50smallest change that would make the path
- 9:52from a raw YouTube idea to a
- 9:54filming-ready script faster. Use the
- 9:56systems already here. Cite the exact
- 9:58files you would reuse and do not edit
- 9:59anything yet." You can see that it
- 10:01mapped the steps from finding an idea
- 10:02through research, to an outline, to a
- 10:04script, and the teleprompter file that I
- 10:06can actually plug in and read while
- 10:08filming. Then it looked at how the work
- 10:09gets passed from one step to the next.
- 10:11The main recommendations about research,
- 10:13my outline skill could gather facts, but
- 10:16it didn't require saving the evidence in
- 10:17a file that script skill knew how to
- 10:19check. And my script skill had a hard
- 10:21rule telling it to verify the facts
- 10:22again. Now, obviously I put that there
- 10:24because I wanted verification. I wanted
- 10:25to make sure that it wasn't making up
- 10:27facts or information. But if we've
- 10:28already found the primary source and
- 10:30checked what it says, I'd like the next
- 10:31step to have that work in front of can
- 10:33see the audit counted 25 video folders
- 10:35with a script, but only three with a
- 10:37file specifically called sources.md.
- 10:39Now, this doesn't mean that all the
- 10:41other videos had no research, but it
- 10:42does show that this particular way of
- 10:44handing over the evidence wasn't
- 10:45consistent. So, it suggested three edits
- 10:47across the two existing skills, and I
- 10:49was able to once again just say, "Yep,
- 10:50that sounds good. Let's go ahead and
- 10:52make those changes." So, now that you
- 10:53understand how you can use Claude to
- 10:54audit your own setup, I would make this
- 10:56something that you're doing regularly.
- 10:58Maybe once a month, you know, if you're
- 10:59adding a bunch of skills and you're
- 11:00changing your workflow all the time, or
- 11:02maybe once a quarter if the setup stays
- 11:04pretty stable. And honestly, what I
- 11:05would recommend is every time a new
- 11:06model drops and you switch into it, you
- 11:08should probably just run this sort of
- 11:09audit anyways, because every model is
- 11:10going to interpret your settings and
- 11:12your skills and your instruction files
- 11:13just a little bit differently. What's
- 11:14really cool is once you've done this a
- 11:15few times and you start to have more
- 11:16trust in the doctor or in like your
- 11:18audit prompts, you can start putting
- 11:20these checks on a schedule. So, for this
- 11:21kind of local setup check, you could use
- 11:23a local scheduled task [music] that can
- 11:25see the same files and settings inside
- 11:26of the Claude desktop app, and you can
- 11:28just leave it open, and then, you know,
- 11:29once once it'll run that doctor, and it
- 11:31will run these different audits, and
- 11:32then you can wake up to recommendations
- 11:34that you can basically just approve or
- 11:36deny. We've talked a lot about AIOS's or
- 11:38AI second brains, and a big pain point
- 11:40with those is on the humans, right?
- 11:42Human consistency is the issue.
- 11:44Sometimes we forget to update things, or
- 11:45we, you know, put duplicate entries in
- 11:47there, or we forget to delete things,
- 11:48stuff like that. So, if you can actually
- 11:50have your AIOS work with you on that
- 11:52cadence of automating the stuff and
- 11:53cleaning the stuff, then you're creating
- 11:54this really powerful, sort of like,
- 11:56self-improving, self-cleaning AI
- 11:58operating system, which is obviously
- 11:59something that there's a lot of value
- 12:01in, because once you have your own
- 12:02personal AIOS figured out, then you can
- 12:04start to figure out how do you have your
- 12:05team stay synced up, and how do you have
- 12:07everyone working together, rather than
- 12:09everyone duplicating work and building
- 12:10their own AIOS's. So, anyways, that is
- 12:12going to do it for this one. I hope that
- 12:13you guys enjoyed or you learned
- 12:14something new, and if you did, please
- 12:16give it a like, it helps me out a ton.
- 12:17And as always, I appreciate you guys
- 12:18making it to the end of the video, and
- 12:20I'll see you on the next one.
- 12:21Thanks, everyone.
About this transcript
This page contains the full transcript of Anthropic Engineers Just 10x'd Everyone's Claude Code by Nate Herk | AI Automation, generated from the public captions YouTube serves with the video. The transcript has 3,151 words across 452 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.
What you can do with it
Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to choose a caption track available on the video.
Free YouTube transcript tool
YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.