YouTube2Text

Anthropic Engineers Just 10x'd Everyone's Claude Code — Transcript

by Nate Herk | AI Automation · 3,151 words · 452 segments · language en · Watch on YouTube

Full transcript

  1. 0:00So, Anthropic engineers just 10x Claude
  2. 0:01code by deleting over 80% of its system
  3. 0:04prompt, but they also changed everything
  4. 0:06that we know about prompting. Because a
  5. 0:07lot of us were told the best way to get
  6. 0:09Claude to behave the way we want it to
  7. 0:11was to give it as much information and
  8. 0:12context as possible. But with this
  9. 0:14change, almost all of your information
  10. 0:16is actually slowing Claude down and
  11. 0:17slowing down its thinking speed. So,
  12. 0:19today I'm going to show you how to fix
  13. 0:21your projects using the engineering
  14. 0:22approach from Anthropic. So, let's get
  15. 0:24into it. All right, so Thoric, one of
  16. 0:26the engineers at Anthropic, recently
  17. 0:27published an article called The New
  18. 0:28Rules of Context Engineering for Claude
  19. 0:305 generation models. His team removed
  20. 0:32over 80% of Claude code system prompt
  21. 0:34for models like Opus 5 and Fable 5 and
  22. 0:365.1 and future models, and they reported
  23. 0:38no measurable loss on their evaluations
  24. 0:40of these models. The system prompt is
  25. 0:42basically the rulebook that comes with
  26. 0:43Claude code. So, your own instructions
  27. 0:45and whatever context gets loaded for the
  28. 0:46task sit right alongside of the system
  29. 0:49prompt. So, they're saying that they
  30. 0:50could cut most of that rulebook and keep
  31. 0:52the performance that they had measured.
  32. 0:53And in my own setup, I used their
  33. 0:55guidance to review the instructions that
  34. 0:56I'd accumulated and clean up how my
  35. 0:58skills work together. And this goes
  36. 0:59right alongside with Boris Cherny, the
  37. 1:01creator of Claude code said himself,
  38. 1:02which was basically like every 6 months
  39. 1:04you should delete all of your skills.
  40. 1:06And we've been told that skills are what
  41. 1:07makes Claude code much smarter and much
  42. 1:09more consistent. Now, Anthropic
  43. 1:10describes that rules they needed for
  44. 1:11older models became too restrictive as
  45. 1:13Claude actually just got better and
  46. 1:15better. One example was comments in code
  47. 1:17because they used to put a hard limit on
  48. 1:18how much Claude could write even though
  49. 1:20some projects needed more explanation.
  50. 1:21[music] The replacement now tells it to
  51. 1:22follow the style of the surrounding
  52. 1:24code. I think it's pretty easy to end up
  53. 1:25with the same problem in your own setup.
  54. 1:27You know, Claude makes a mistake, you
  55. 1:29add a rule, and then a few months later
  56. 1:30you've got that rule in your Claude.md
  57. 1:32file or in a skill, maybe a slightly
  58. 1:34different version of it somewhere else,
  59. 1:35and you don't actually need that rule
  60. 1:37because you keep upgrading the model,
  61. 1:38but the instructions still stay in
  62. 1:40there. And the model might not need all
  63. 1:41of that extra context now. I recently
  64. 1:43did a video about prompting Fable 5.1. I
  65. 1:45talked about the idea of defining to the
  66. 1:47model what done looks like. You
  67. 1:49basically need to give it the outcome
  68. 1:50that you're after, why you're doing it,
  69. 1:52and the constraints that it needs to
  70. 1:53respect. And then you just have to get
  71. 1:54out of its way. Leave it some room to
  72. 1:56actually do [music] the work. So, for
  73. 1:57this video, I asked Claude to look at
  74. 1:58how I get from a YouTube idea to a
  75. 2:00script that I could film [music] or an
  76. 2:01outline that I could film about, and I
  77. 2:02wanted it to find a small improvement
  78. 2:04and point me to the file so that I could
  79. 2:05check the recommendation for myself. So
  80. 2:07I'll show you guys that demo of me
  81. 2:08actually doing this later in the video.
  82. 2:10Now, Anthropic's applied AI team
  83. 2:11recommends using the smallest amount of
  84. 2:13context that gets the result you want,
  85. 2:15and I turn that into sort of my own
  86. 2:16saying, which is the best context is the
  87. 2:19smallest amount of context that gets you
  88. 2:20the highest quality result at the
  89. 2:22cheapest cost. They also explicitly say
  90. 2:24that minimal doesn't necessarily mean
  91. 2:26short. If the model needs a detail to do
  92. 2:28the job, then obviously you need to keep
  93. 2:29that detail in there. And there's
  94. 2:30research looking at what happens when
  95. 2:32you add more input. A paper published in
  96. 2:342025 tested five models on math,
  97. 2:36question answering, and coding, and
  98. 2:38performance dropped as the input got
  99. 2:39longer, even when the models could
  100. 2:41retrieve the relevant information. The
  101. 2:43reported drops ranged from 13.9% to 85%
  102. 2:46depending on the model and the task.
  103. 2:48Now, those were obviously different
  104. 2:49models, and that was, you know, over a
  105. 2:50year and a half ago, and these were
  106. 2:52controlled experiments, so those
  107. 2:53percentages don't tell us what will
  108. 2:54happen when you trim your own Claude 3
  109. 2:56MD, for example. But Anthropic also
  110. 2:57tested loading MCP tool definitions only
  111. 2:59when needed. They reported an 85%
  112. 3:02reduction in token usage in their
  113. 3:03example, and their MCP evaluation scores
  114. 3:05improved from 49% to 74% for Opus 4, and
  115. 3:09from 79.5 to 88.1% for Opus 4.5. So
  116. 3:13there are measured results behind this
  117. 3:14exact idea. And in that case, Claude
  118. 3:16still had access to the tools. It just
  119. 3:18didn't have to load all of those
  120. 3:20instructions up front. Now, in your own
  121. 3:21setup, you can use a slash doctor, and
  122. 3:23when I recently did this, the doctor
  123. 3:24report gave me this workflow audit with
  124. 3:26specific instructions and handoffs that
  125. 3:28I need to clean up. So I'm also going to
  126. 3:29show you guys that in just a sec. Real
  127. 3:30quick, guys, a message from today's
  128. 3:31sponsor, HyperAgent. If you're building
  129. 3:33agents for clients, most of the work
  130. 3:35lands around the agent. Because each
  131. 3:37client's data has to stay separate,
  132. 3:39their staff needs to use it without
  133. 3:40breaking it, and somebody has to be
  134. 3:42obviously paying for those runs. So what
  135. 3:44HyperAgent does is they give every
  136. 3:46client their own workspace. So the
  137. 3:47agents, the skills, the memories, and
  138. 3:49the documents all live inside one
  139. 3:51perimeter, so nothing from one client
  140. 3:53shows up in the, you know, the context
  141. 3:55of another agent. And your builders work
  142. 3:56in there with full edit access, and the
  143. 3:58client's own staff can get a role that
  144. 4:00runs the agents, but can't touch the
  145. 4:01system prompts or swap the models or
  146. 4:03accidentally delete a skill or anything
  147. 4:05like that. So, if that sounds useful for
  148. 4:07your agency, then check out Hyper Agent
  149. 4:08through the link in the description.
  150. 4:10Huge thanks to Hyper Agent for
  151. 4:11sponsoring this part of the video. Now,
  152. 4:12let's get back to it. Now, I have tried
  153. 4:13some pretty extreme versions of this,
  154. 4:15just experimenting with my own AI
  155. 4:16operating system. You know, I've
  156. 4:17duplicated it before, I've removed the
  157. 4:19Claude 3 MD I've removed my skills, and
  158. 4:20I've tried off, you know, a bunch of
  159. 4:22different prompts to see them do the
  160. 4:23same job. So, one example that I
  161. 4:24recently did was I took this interview
  162. 4:26with Boris Cherney, and I wanted to turn
  163. 4:28it into a student resource guide. Now,
  164. 4:29my full AIOS setup that I was used to
  165. 4:31running every day, it made the prettier
  166. 4:33document because it had my branding, it
  167. 4:35had a proper heading, it had links that
  168. 4:36I normally include, and it had just like
  169. 4:38more specific context about me, all that
  170. 4:39kind of stuff. And the fresh version
  171. 4:41looked a little bit messier, it wasn't
  172. 4:42as pretty, but its structure was much
  173. 4:44better. It had better details, it
  174. 4:46organized the interview into ideas, and
  175. 4:47it added timestamps, and it just felt
  176. 4:49more professional, and it felt like a
  177. 4:50better guide to actually give to a
  178. 4:52student. So, what I had was a polished
  179. 4:53guide from one version and a content
  180. 4:55structure that I preferred from the
  181. 4:57other version. Now, obviously, I still
  182. 4:58want my branding, so what I would do is
  183. 5:00I would build a new skill out of this,
  184. 5:01and I would say, "Hey, you know, you
  185. 5:02should be using this as the header and
  186. 5:04this as the font and this as the footer
  187. 5:05and add these logos and things like
  188. 5:06that. But, I'm going to leave it up to
  189. 5:08you to actually structure the content in
  190. 5:09the way that you see fit because I've
  191. 5:11gave you the basically the intent of
  192. 5:13turn this into a student resource
  193. 5:14guide." Because in the pretty version
  194. 5:16that had my skill already, I had so many
  195. 5:18guardrails and so many instructions that
  196. 5:20I was basically just like confining the
  197. 5:21model, and I wasn't letting it truly be
  198. 5:23creative and give me the output that it
  199. 5:24thought was best because it was reading
  200. 5:26my skill, and it had to follow my
  201. 5:27instructions. And those instructions
  202. 5:29were helpful back when we had like Opus
  203. 5:314, but now that we have Fable 5.1, it's
  204. 5:33able to just do a better job on its own.
  205. 5:35It's kind of like giving an experienced
  206. 5:36designer the checklist that you would
  207. 5:37give to a kid making their first
  208. 5:39presentation because the designer still
  209. 5:40needs to know who it's for and what
  210. 5:41you're trying to accomplish. But, if you
  211. 5:43dictate the placement of every box and
  212. 5:45every sentence before they've even seen
  213. 5:47the material, you might prevent them
  214. 5:48from coming up with something that they
  215. 5:50might add that you would actually like
  216. 5:51more. And that's how I think about
  217. 5:52reviewing these instructions. I want to
  218. 5:54keep the details that help Claude
  219. 5:55understand the job and what makes it
  220. 5:57specific to me, and then test it whether
  221. 5:59it still needs the procedure that I've
  222. 6:00written around it. So, when I talk about
  223. 6:01the context part of my four C's of an AI
  224. 6:04OS, this is the kind of information I
  225. 6:06mean. I mean my voice, I mean my goals,
  226. 6:08I mean my audience, and the way that my
  227. 6:09business works. Claude can't just infer
  228. 6:11all of that or look it up online, so
  229. 6:13that's what I have to supplement to it.
  230. 6:15Now, it's also important to explain why
  231. 6:17you're asking for something, the intent
  232. 6:18behind it. So, if I'm making a resource
  233. 6:20guide for students, that helps Claude
  234. 6:22make decisions about using, you know,
  235. 6:24analogies and simple explanations and
  236. 6:25using examples. And I'd keep that
  237. 6:27context when I clean up the skill. But,
  238. 6:29a long folder listing is something it
  239. 6:30can probably look up. So, if the same
  240. 6:32instruction is repeated in multiple
  241. 6:34different places, I'd check whether I
  242. 6:36need all of those copies or if I could
  243. 6:37just clean them up. And if a detailed
  244. 6:39procedure only applies to when I'm
  245. 6:41editing a video, I'd rather load it in
  246. 6:43when I'm actually editing a video. And
  247. 6:44this is all about getting out of the
  248. 6:46model's way, which still means you have
  249. 6:47to be clear about what it's allowed to
  250. 6:49do. So, I'd keep the Claude on MD useful
  251. 6:51as a starting point. Obviously, telling
  252. 6:52Claude things like what the project is
  253. 6:53for, giving it any of the gotchas and
  254. 6:55the landmines to watch out for, and
  255. 6:57point it toward the deeper instructions.
  256. 6:59So, now my whole video editing process
  257. 7:01doesn't need to be there when I'm asking
  258. 7:02for help with a script. So, inside
  259. 7:04Claude code, you can run {slash} doctor
  260. 7:06to look through a lot of the stuff
  261. 7:07happening locally that you probably
  262. 7:08don't spend much time thinking about,
  263. 7:10settings, config files, tedious stuff
  264. 7:12like that. So, I ran {slash} doctor in a
  265. 7:14new session, and it checked my skills,
  266. 7:16my memory, my plugins, my settings,
  267. 7:18[music] and my installation. And you can
  268. 7:19see in this table right here, it lists
  269. 7:21each skill, where it's installed, and
  270. 7:22how many recorded uses it has since
  271. 7:24installation, and an estimate of the
  272. 7:26tokens that it's listing takes up. That
  273. 7:28basically just means like the YAML front
  274. 7:30matter. It uses that YAML front matter
  275. 7:32to find the skill, which is different
  276. 7:34from loading in the full instruction
  277. 7:36when it actually decides to run it and
  278. 7:37read the whole thing. For example, Agent
  279. 7:39Builder is estimated at 111 tokens, and
  280. 7:42Prove it is 103 tokens. And both show
  281. 7:44zero recorded uses here. The report
  282. 7:46flagged 18 older skills with no recorded
  283. 7:48use, but kept the ones that I'd only
  284. 7:50created a few days ago. So, it's taking
  285. 7:52their age into account as well. It also
  286. 7:53checks 50 recent sessions from this
  287. 7:55project and says that it also looked at
  288. 7:57lifetime counters across 252 startups
  289. 7:59and skill and plug-in dispatches across
  290. 8:02223 transcripts. So, I'd still review
  291. 8:04that list before I decide to turn
  292. 8:05anything off or clean anything up or
  293. 8:06delete anything, especially if I do use
  294. 8:08a skill through a different tool for
  295. 8:09example. But, these are the usage
  296. 8:11records available to Claude code in this
  297. 8:13audit. And then it also checks for
  298. 8:14things like duplication. So, in mine,
  299. 8:16the project map was repeating things
  300. 8:18already covered by the routing map.
  301. 8:19There were also repeated sections for
  302. 8:21templates, references, and archiving
  303. 8:22files, and it identified about 2,250
  304. 8:25characters of cuts with an estimated
  305. 8:27saving of 563 tokens. And you can see it
  306. 8:29breaks out different places that memory
  307. 8:31is actually coming from. You know, I've
  308. 8:32got my project level Claude MD, my
  309. 8:34user-wide instructions, the auto memory
  310. 8:36index, and my Claude local MD for
  311. 8:39instructions specific to this checkout
  312. 8:41on my machine. The local file was just
  313. 8:43an empty template, estimated at about 35
  314. 8:45tokens. Now, that's obviously not a ton,
  315. 8:46but the point I'm trying to make here is
  316. 8:47that it will look through a lot of the
  317. 8:49stuff that you probably don't even think
  318. 8:50about ever. It also found the Claude MD
  319. 8:52from the separate video project in my
  320. 8:53desktop folder that loads into Herc too,
  321. 8:56which is it roughly another 2,000 tokens
  322. 8:58just by its estimate here. So, even if
  323. 8:59you've cleaned up the file inside your
  324. 9:01project, there can be instructions
  325. 9:02coming from somewhere else that you
  326. 9:03don't even know you're technically
  327. 9:05paying for. So, the setup checks caught
  328. 9:06two broken skills as well. One had the
  329. 9:08wrong file name, the other had a
  330. 9:10formatting problem in its front matter,
  331. 9:11the little description block at the top
  332. 9:13like we talked about. So, Claude wasn't
  333. 9:14actually properly seeing its
  334. 9:15description. And those files have been
  335. 9:17corrected because at the end of the
  336. 9:18doctor audit, it says, "Hey, do you want
  337. 9:20me to just fix all of this?" It also
  338. 9:21looks at things like your installation,
  339. 9:23it looks at things like your automatic
  340. 9:24updates, and it checked the settings
  341. 9:25behind auto mode, permissions, and my
  342. 9:27hooks. And real quick guys, if you're
  343. 9:29learning how to build with AI, I've got
  344. 9:30a free community with a bunch of
  345. 9:31resources and skills and courses on
  346. 9:33building AI systems and agents, and you
  347. 9:35can also ask questions and work through
  348. 9:36the stuff with other people who are
  349. 9:38also, you know, behind you or in front
  350. 9:39of you. So, if you want to join, the
  351. 9:40link is in the description. Let's get
  352. 9:42back to the video. All right. So, after
  353. 9:43running /doctor, I asked Claude Code to
  354. 9:45actually look at my YouTube workflow.
  355. 9:47And this is the prompt that I used. I
  356. 9:48said, "Audit this repo and show me the
  357. 9:50smallest change that would make the path
  358. 9:52from a raw YouTube idea to a
  359. 9:54filming-ready script faster. Use the
  360. 9:56systems already here. Cite the exact
  361. 9:58files you would reuse and do not edit
  362. 9:59anything yet." You can see that it
  363. 10:01mapped the steps from finding an idea
  364. 10:02through research, to an outline, to a
  365. 10:04script, and the teleprompter file that I
  366. 10:06can actually plug in and read while
  367. 10:08filming. Then it looked at how the work
  368. 10:09gets passed from one step to the next.
  369. 10:11The main recommendations about research,
  370. 10:13my outline skill could gather facts, but
  371. 10:16it didn't require saving the evidence in
  372. 10:17a file that script skill knew how to
  373. 10:19check. And my script skill had a hard
  374. 10:21rule telling it to verify the facts
  375. 10:22again. Now, obviously I put that there
  376. 10:24because I wanted verification. I wanted
  377. 10:25to make sure that it wasn't making up
  378. 10:27facts or information. But if we've
  379. 10:28already found the primary source and
  380. 10:30checked what it says, I'd like the next
  381. 10:31step to have that work in front of can
  382. 10:33see the audit counted 25 video folders
  383. 10:35with a script, but only three with a
  384. 10:37file specifically called sources.md.
  385. 10:39Now, this doesn't mean that all the
  386. 10:41other videos had no research, but it
  387. 10:42does show that this particular way of
  388. 10:44handing over the evidence wasn't
  389. 10:45consistent. So, it suggested three edits
  390. 10:47across the two existing skills, and I
  391. 10:49was able to once again just say, "Yep,
  392. 10:50that sounds good. Let's go ahead and
  393. 10:52make those changes." So, now that you
  394. 10:53understand how you can use Claude to
  395. 10:54audit your own setup, I would make this
  396. 10:56something that you're doing regularly.
  397. 10:58Maybe once a month, you know, if you're
  398. 10:59adding a bunch of skills and you're
  399. 11:00changing your workflow all the time, or
  400. 11:02maybe once a quarter if the setup stays
  401. 11:04pretty stable. And honestly, what I
  402. 11:05would recommend is every time a new
  403. 11:06model drops and you switch into it, you
  404. 11:08should probably just run this sort of
  405. 11:09audit anyways, because every model is
  406. 11:10going to interpret your settings and
  407. 11:12your skills and your instruction files
  408. 11:13just a little bit differently. What's
  409. 11:14really cool is once you've done this a
  410. 11:15few times and you start to have more
  411. 11:16trust in the doctor or in like your
  412. 11:18audit prompts, you can start putting
  413. 11:20these checks on a schedule. So, for this
  414. 11:21kind of local setup check, you could use
  415. 11:23a local scheduled task [music] that can
  416. 11:25see the same files and settings inside
  417. 11:26of the Claude desktop app, and you can
  418. 11:28just leave it open, and then, you know,
  419. 11:29once once it'll run that doctor, and it
  420. 11:31will run these different audits, and
  421. 11:32then you can wake up to recommendations
  422. 11:34that you can basically just approve or
  423. 11:36deny. We've talked a lot about AIOS's or
  424. 11:38AI second brains, and a big pain point
  425. 11:40with those is on the humans, right?
  426. 11:42Human consistency is the issue.
  427. 11:44Sometimes we forget to update things, or
  428. 11:45we, you know, put duplicate entries in
  429. 11:47there, or we forget to delete things,
  430. 11:48stuff like that. So, if you can actually
  431. 11:50have your AIOS work with you on that
  432. 11:52cadence of automating the stuff and
  433. 11:53cleaning the stuff, then you're creating
  434. 11:54this really powerful, sort of like,
  435. 11:56self-improving, self-cleaning AI
  436. 11:58operating system, which is obviously
  437. 11:59something that there's a lot of value
  438. 12:01in, because once you have your own
  439. 12:02personal AIOS figured out, then you can
  440. 12:04start to figure out how do you have your
  441. 12:05team stay synced up, and how do you have
  442. 12:07everyone working together, rather than
  443. 12:09everyone duplicating work and building
  444. 12:10their own AIOS's. So, anyways, that is
  445. 12:12going to do it for this one. I hope that
  446. 12:13you guys enjoyed or you learned
  447. 12:14something new, and if you did, please
  448. 12:16give it a like, it helps me out a ton.
  449. 12:17And as always, I appreciate you guys
  450. 12:18making it to the end of the video, and
  451. 12:20I'll see you on the next one.
  452. 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.