YouTube2Text

So nutzt du Claude Code nach Anthropics neuen Regeln — Transcript

by Julian Ivanov | KI-Automatisierung · 5,623 words · 833 segments · language en · Watch on YouTube

Full transcript

  1. 0:00This summer, Anthropic deleted about 80
  2. 0:02%of Claude's system prompt—the
  3. 0:04instructions that Claude automatically
  4. 0:06receives with every single message—
  5. 0:08and according to Anthropic, the model
  6. 0:10has actually become better because of
  7. 0:12it. At the same time, the documentation
  8. 0:15was rewritten in quite a few places,
  9. 0:16because with Opus 3.5 and Sonnet 3.5,
  10. 0:18the models behave differently than they
  11. 0:20did just half a year ago. This means
  12. 0:22that some of the habits you've picked
  13. 0:24up might not be entirely optimal as of
  14. 0:26today and could actually be slowing
  15. 0:28Claude down. That's why today I'm going
  16. 0:31to go through 12 mistakes that many
  17. 0:32people still make when working with
  18. 0:34Claude, and for each one, I'll show you
  19. 0:36what you should do instead. Claude
  20. 0:37Anthropic. Let's start from the very
  21. 0:39beginning with the pure prompting of
  22. 0:41these new models. The models have
  23. 0:44become so good by now that we also need
  24. 0:46to adjust the way we prompt to really
  25. 0:48get the quality these models can
  26. 0:49deliver. This applies not just to
  27. 0:52Claude, but also here in the normal
  28. 0:54Claude chat window or in Projects.
  29. 0:56Let's look at a concrete example.
  30. 0:58Assume I want to write a newsletter for
  31. 1:00my community and I want Claude to do it
  32. 1:02for me. Many people write the prompt in
  33. 1:05roughly this style: "You are an
  34. 1:06excellent newsletter author with years
  35. 1:08of experience. Write me this week's
  36. 1:11newsletter on the topic of new best
  37. 1:12practices for Claude. Write in my style
  38. 1:15, keep it casual, and not too long."
  39. 1:16Even the top models will still have
  40. 1:18trouble with a prompt like that and
  41. 1:20won't achieve the result you might be
  42. 1:22imagining. Just look at everything the
  43. 1:24model has to guess with this prompt.
  44. 1:26First, who the newsletter is for and
  45. 1:27what the readers should do with it.
  46. 1:29Then, what does "my style" even mean,
  47. 1:31and what does "not too long" mean? We
  48. 1:34also aren't giving it any reference
  49. 1:36material here that would give Claude a
  50. 1:37clue about how we like to write
  51. 1:39newsletters. Additionally, it has been
  52. 1:41discovered that phrases like "you are
  53. 1:43an excellent newsletter author with
  54. 1:45years of experience" do nothing to help
  55. 1:47the model provide better output. Things
  56. 1:50like that are actually ignored by the
  57. 1:51model, and the only thing it takes from
  58. 1:53that role is simply the style. So, if I
  59. 1:56say "you are a Romantic-era author,
  60. 1:57write the following text," the model is
  61. 1:59already trying its best to give you the
  62. 2:02best possible output if you just say
  63. 2:03you need a newsletter. It doesn’t
  64. 2:05help anymore if you say, yes, this is
  65. 2:07extremely important or I need your full
  66. 2:08concentration or something like that.
  67. 2:11Anthropic has created a prompt library
  68. 2:13in their documentation with a section
  69. 2:14on what actually makes a good prompt.
  70. 2:17And you can really break the whole
  71. 2:19thing down into three essential points.
  72. 2:20And the first is that you should never
  73. 2:22just give the pure instruction or
  74. 2:24request to the model, but always also
  75. 2:26the reason for it and your goal. The
  76. 2:29documentation says the model works
  77. 2:31better if it knows the intention behind
  78. 2:33the task, because then it can connect
  79. 2:34the task to what else it knows, instead
  80. 2:36of guessing your intention. And the
  81. 2:40template there effectively says: I am
  82. 2:42currently working on this task for this
  83. 2:44target audience. They need the
  84. 2:46following output, and with this
  85. 2:48information, please carry out the
  86. 2:49following task. Second, it is also
  87. 2:52usually better to show Claude something
  88. 2:53instead of describing it. If I write
  89. 2:55something to Claude like I write in a
  90. 2:57casual style and not too long, then it
  91. 2:59is incredibly difficult for Claude to
  92. 3:01replicate my style if I haven't defined
  93. 3:03it anywhere in a file. When you work in
  94. 3:05Claude Code, you can use an@to
  95. 3:07reference specific files that Claude
  96. 3:08Code then reads directly along with the
  97. 3:10prompt. And here, for example, it would
  98. 3:13make a lot of sense to simply reference
  99. 3:15past newsletters that I have written
  100. 3:16myself. Claude can then read through
  101. 3:18them, and then this newsletter will
  102. 3:20have the new writing style. That will
  103. 3:21work significantly better than if you
  104. 3:23try to describe your way of writing to
  105. 3:25it somehow. And the third point, which
  106. 3:27is also extremely important, is what
  107. 3:28does finished even mean and how can
  108. 3:30Claude check the result itself?
  109. 3:31According to the documentation, you
  110. 3:33describe the result with a measurable
  111. 3:35goal so that it is clear when Claude
  112. 3:37can stop. So, if I now write something
  113. 3:39like the newsletter shouldn't be too
  114. 3:40long, that’s not a real number. But
  115. 3:42if I say 500 words, three to four
  116. 3:44paragraphs, and a small tip at the end,
  117. 3:46then that is much more measurable, and
  118. 3:48Claude can do much more with it, and
  119. 3:50then you give Claude the opportunity to
  120. 3:52check its result itself. This is the
  121. 3:54first essential best practice in the
  122. 3:56entire best practice documentation here
  123. 3:58from Claude Code, because the models
  124. 3:59have become much better at working on a
  125. 4:01task for a long time. This means Claude
  126. 4:03Code can stay in this agentic loop
  127. 4:05until the result is exactly how you
  128. 4:07want it to be. If you are just writing
  129. 4:09a longer text now, you can also say
  130. 4:11something like: "Look at the past
  131. 4:13newsletters now." e.g., the ones we
  132. 4:15wrote, at the end, and check if the new
  133. 4:17newsletter you generated matches the
  134. 4:19old ones—in tone, length, content,
  135. 4:22and so on. This means it’s not enough
  136. 4:24to just tell it: "Yes, check yourself
  137. 4:25at the end, check your result," but you
  138. 4:27must also tell it what to check against
  139. 4:29—what are the evaluation criteria?
  140. 4:30And last but not least, it can be very
  141. 4:32useful to tell the model what you do
  142. 4:34not expect in the result or what you do
  143. 4:36not want it to do. Because especially
  144. 4:38Opus 3.5 and Sonnet 3.5 often tend to
  145. 4:40come up with additional tasks that you
  146. 4:42didn't assign at the beginning, and
  147. 4:44then they do something for you, and you
  148. 4:45think, why did you just do that? I
  149. 4:48didn't even ask for that, and how did
  150. 4:50you even figure out you could do that,
  151. 4:52and so on. It even says literally in
  152. 4:54the documentation here for Opus 3.5,
  153. 4:56for example, because the model wants to
  154. 4:58take everything off your hands as much
  155. 4:59as possible, but there are certain
  156. 5:01things you explicitly don't want. And
  157. 5:03if you know that, if you know what you
  158. 5:05don't want in the result, then you
  159. 5:07should clearly write that in the prompt
  160. 5:08, otherwise the model will just do it
  161. 5:10because it thinks it would be useful
  162. 5:11for you. That means a much better
  163. 5:13prompt than the one at the beginning
  164. 5:14would be something like this. I write a
  165. 5:16newsletter for my community every week.
  166. 5:18They are primarily self-employed
  167. 5:19entrepreneurs and AI enthusiasts. Many
  168. 5:21are already working with Claude. Every
  169. 5:23week, they should take away a few best
  170. 5:24practices and news items that they can
  171. 5:26implement directly in their work with
  172. 5:27AI. With that, we immediately have the
  173. 5:29background, the target group, and what
  174. 5:31we want to achieve. Then I give it the
  175. 5:33reference. The topic this week is the
  176. 5:35new best practices for Claude. My notes
  177. 5:38on this are in this file. You can't
  178. 5:41reference that directly here in the
  179. 5:42Claude chat. That works in Claude for
  180. 5:44developers. And for the style and
  181. 5:46structure, take the last three issues
  182. 5:47—and then I give it the folder with
  183. 5:49my last newsletters, for example. At
  184. 5:51the end, I also say what "finished"
  185. 5:52means to me: namely, when the topic is
  186. 5:54explained in three to four paragraphs
  187. 5:56and there is one thing at the end that
  188. 5:58you can try out this week. Compare the
  189. 6:00tone and length at the end with the
  190. 6:02three issues, and I only want the
  191. 6:03newsletter text, no subject line or any
  192. 6:05other things. So, at the beginning, we
  193. 6:07have the background and the goal. We
  194. 6:09have references that Claude can use,
  195. 6:11and we define what our final result
  196. 6:12should look like and how it can verify
  197. 6:14it. Believe me, if you follow these
  198. 6:17three things when working with agentic
  199. 6:18systems like Claude, Hermes, Codex, and
  200. 6:20so on, you will get significantly
  201. 6:22better results. And since this is a
  202. 6:25task I do every week, it would be very
  203. 6:27smart to turn it into a skill once
  204. 6:29Claude has done it with me and we are
  205. 6:31happy with the result. That is then
  206. 6:34essentially a set of instructions that
  207. 6:35Claude can use every time I want to
  208. 6:36write a newsletter. I can then call up
  209. 6:38this skill every week when I write a
  210. 6:40newsletter, and Claude immediately
  211. 6:42knows the drill and what it needs to do
  212. 6:44, and I can additionally provide it
  213. 6:45with references for the latest topics
  214. 6:47and so on. Let's move on to the second
  215. 6:49mistake, which is that many people
  216. 6:51still switch models in the same chat
  217. 6:53here and there while chatting with
  218. 6:55Claude to save a few tokens. Claude
  219. 6:57sends the entire previous chat history
  220. 7:00with every message. But so that this
  221. 7:02doesn't fully impact your quota every
  222. 7:03time, there is a so-called cache. This
  223. 7:05means that what has already been
  224. 7:07processed only counts as a fraction the
  225. 7:09next time. And this cache is tied to
  226. 7:11the specific model you are currently
  227. 7:12using. However, if you switch models in
  228. 7:15the middle of a chat, the entire chat
  229. 7:17history is re-uploaded to the model,
  230. 7:19and all those tokens are consumed once
  231. 7:21again. And that naturally eats into
  232. 7:24your quota, because all those tokens
  233. 7:25have to be fully processed again. And
  234. 7:27that is the reason why some people hit
  235. 7:29their session limit after just two
  236. 7:30hours, even though they haven't done as
  237. 7:32much as they claim. This means you
  238. 7:34should avoid switching back and forth
  239. 7:35between models too much. It is best to
  240. 7:38work with one model on one task, finish
  241. 7:40it, and if you realize you want to do
  242. 7:42something else, start a new task, then
  243. 7:44start a new chat with that specific
  244. 7:46model, be it Opus, Sonnet, or Haiku.
  245. 7:49Let's come to the third mistake, which
  246. 7:51is telling Claude too often what it
  247. 7:53should not do. That was a big point
  248. 7:55that Anthropic also shortened in the
  249. 7:57Claude documentation for Claude,
  250. 7:59because they had specific instructions
  251. 8:01in many places about what the model
  252. 8:03should not do. And the problem is, if
  253. 8:05you keep telling it what not to do,
  254. 8:07you'll confuse the model in the long
  255. 8:09run, because the more prohibitions you
  256. 8:11define, the more it has to weigh each
  257. 8:12individual prohibition against the rest
  258. 8:14of the instructions. And Anthropic has
  259. 8:16realized that it makes more sense,
  260. 8:18instead of giving the model a lot of
  261. 8:19prohibitions, to rather say exactly
  262. 8:21what it should do concretely. That
  263. 8:23means, instead of saying don't be too
  264. 8:24long-winded, it is better to write,
  265. 8:26answer in no more than five sentences.
  266. 8:28Or instead of writing no bullet points,
  267. 8:29I prefer to write, write in full
  268. 8:31sentences. And if you have very
  269. 8:32specific things that it really
  270. 8:34shouldn't do, then that's no problem
  271. 8:35either, you can of course tell it that.
  272. 8:37But the point here is really just that
  273. 8:38you should rather tell Claude what it
  274. 8:40should do, instead of what it should
  275. 8:42not do, so that you simply avoid such a
  276. 8:43list of prohibitions, which at some
  277. 8:45point will contradict each other and
  278. 8:47just confuse Claude, because it then
  279. 8:48has to deduce what it should do instead
  280. 8:50. And with that, we come to the fourth
  281. 8:52mistake. If we want Claude to really
  282. 8:54always do something specific, most
  283. 8:56people just write another rule in the
  284. 8:58Claude.md file along the lines of
  285. 9:00please do it this way every time. But
  286. 9:03the problem with that is that this
  287. 9:04Claude.md is ultimately just additional
  288. 9:06context for Claude. Just like the
  289. 9:08system prompt, which is sort of already
  290. 9:09built into Claude's code by Anthropic.
  291. 9:11That is also always loaded into the
  292. 9:12context. Just like the tools and skills
  293. 9:14that Claude can use. Just like the
  294. 9:16entire chat history and also the files
  295. 9:18that Claude has read. All this
  296. 9:20information is loaded into Claude's
  297. 9:22context, and the longer you work on the
  298. 9:24task and the more the context window
  299. 9:26fills up, the more often it can happen
  300. 9:28that instructions are simply not
  301. 9:29followed because they get lost under
  302. 9:31all that context. This means the
  303. 9:33Claude.md file is not a set of rules
  304. 9:34that always applies and that Claude
  305. 9:36always follows. If I write in my
  306. 9:38Claude.md, for example, always write
  307. 9:39German texts with umlauts, then that
  308. 9:41will happen most of the time, but there
  309. 9:43will always be cases where it forgets.
  310. 9:45Especially when I've been working on
  311. 9:46the task for a longer time. But if you
  312. 9:47really want to ensure that a certain
  313. 9:49action is always carried out, then you
  314. 9:51should definitely use hooks. Hooks
  315. 9:53automatically execute scripts at
  316. 9:55certain points in Claude's workflow.
  317. 9:57Unlike the instructions in the
  318. 9:59Claude.md, which are merely
  319. 10:01recommendations, hooks are
  320. 10:02deterministic and guarantee the
  321. 10:04execution of the action. So, if I want
  322. 10:07to ensure that Claude......for example,
  323. 10:09when creating files and including
  324. 10:11German text, actually writes umlauts
  325. 10:13correctly, I create a hook for that,
  326. 10:14which then deterministically checks
  327. 10:16every time that it happens. And if not,
  328. 10:19then Claude is notified and has to
  329. 10:21rewrite the text. If I want to ensure,
  330. 10:24for instance, that Claude doesn't look
  331. 10:25at certain files because they contain
  332. 10:27credentials or other data that Claude
  333. 10:29shouldn't read, I can set up a hook for
  334. 10:31that as well. Because as soon as Claude
  335. 10:34tries to read the file, this hook is
  336. 10:35activated and blocks the action. And if
  337. 10:38you are wondering now how to create
  338. 10:39such a hook, you just have to tell
  339. 10:40Claude Code. Just tell it: "Hey, I
  340. 10:42would like you to create a hook that
  341. 10:44ensures that this and that always
  342. 10:45happens in such a way." This really
  343. 10:48ensures that it happens, and you don't
  344. 10:50have to rely on Claude having
  345. 10:51remembered it from the instructions
  346. 10:53within the whole context it has to
  347. 10:54process in parallel. And let's stick
  348. 10:57with the Claude.md file, because many
  349. 10:58people still make mistakes here. And
  350. 11:00the fifth mistake is changing the
  351. 11:01Claude.md file in the middle of a
  352. 11:03session. If you are working on a task
  353. 11:05with Claude Code and you notice that
  354. 11:07Claude isn't doing something right—it
  355. 11:09keeps getting it wrong—and you want
  356. 11:11to add that to the Claude.md, you add
  357. 11:13it, but then continue in the chat, the
  358. 11:15new version of this Claude.md file will
  359. 11:17not be loaded. This file is only ever
  360. 11:20loaded once at the beginning of the
  361. 11:22session. That means any changes made
  362. 11:23afterward are not taken into account in
  363. 11:25this session. This leads to some people
  364. 11:27complaining that Claude isn't following
  365. 11:28the instructions despite them being in
  366. 11:30the file. But that is because you
  367. 11:31simply have to start a new session so
  368. 11:33that the new Claude.md file is taken
  369. 11:35into account. This brings us to the
  370. 11:37sixth mistake, which is that people
  371. 11:38still underestimate how important it is
  372. 11:40to keep the context window of these
  373. 11:42language models as empty as possible.
  374. 11:45The more information the language model
  375. 11:46has to process at the same time and the
  376. 11:48more the context window is filled, the
  377. 11:50dumber it becomes and the worse it gets
  378. 11:52at retrieving specific information. It
  379. 11:54simply loses track. Anthropic measures
  380. 11:56this with a benchmark called Needle In
  381. 11:58A Haystack. There, various pieces of
  382. 12:00information are hidden in a huge text,
  383. 12:02and then it is tested how many of them
  384. 12:03the model can find again. And by now,
  385. 12:05the Claude models have a context window
  386. 12:07of one million tokens. That means it
  387. 12:09can process 1 million tokens, or about
  388. 12:12750,000 words, at the same time. But
  389. 12:14that doesn't mean the model can
  390. 12:16actually retrieve every piece of
  391. 12:17information, because at the end of the
  392. 12:19day, it's all based on probabilities.
  393. 12:21And it's been shown that when the model
  394. 12:24processes only about 256,000 tokens, it
  395. 12:26finds 93%of the requested information,
  396. 12:29but as soon as it processes 1 million
  397. 12:31tokens, that drops to 76%. This means
  398. 12:36in a packed chat, when the context
  399. 12:37window is completely full, it can
  400. 12:39happen that on average every fourth
  401. 12:41piece of information is lost or Claude
  402. 12:43simply loses track of it, and that
  403. 12:45includes your rules and instructions.
  404. 12:48So it's really worth regularly starting
  405. 12:50a new chat, clearing the context, or
  406. 12:53using the slash-compact command to
  407. 12:55summarize the conversation history and
  408. 12:57the entire context to continue working.
  409. 13:00Because just take a look at this
  410. 13:01context window simulation from Claude
  411. 13:03Code here. Here the documentation shows
  412. 13:05how much context Claude Code actually
  413. 13:07processes with every session we start.
  414. 13:09And as soon as I start the session now,
  415. 13:11the following things are loaded into
  416. 13:13the language model's context window.
  417. 13:15System prompt, auto-memory, environment
  418. 13:17info, MCP tools, skills, the Claude.md,
  419. 13:19and all that before I've even written
  420. 13:21anything. And when I then write a
  421. 13:23message, like "fix this bug," my prompt
  422. 13:26is added here, but also every file that
  423. 13:29Claude reads through. That means just
  424. 13:31after this one message, I've already
  425. 13:33used up 21,000 tokens, and people
  426. 13:34underestimate how fast that happens. By
  427. 13:37the way, you can type/context in the
  428. 13:38chat with Claude Code at any time to
  429. 13:40see how the context window is currently
  430. 13:42filled, which often shows you what you
  431. 13:44might be able to optimize. So, really
  432. 13:47get into the habit of starting a new
  433. 13:49chat or using/clear when you're done
  434. 13:51with a task, or using something like/
  435. 13:53compact. Then you keep a summary of
  436. 13:55what was done before. You can also
  437. 13:57define what you want to have summarized
  438. 13:59. Or what you can also do—many people
  439. 14:01do this—is have a handoff written for
  440. 14:03the current session. That just means a
  441. 14:06summary in a file that Claude can then
  442. 14:08read when you start a new session. This
  443. 14:10is also a great skill to have, right?
  444. 14:12Then next time you can just write/
  445. 14:14handoff, for example, and Claude will
  446. 14:15create it automatically. Mad Pocock,
  447. 14:18who has a really great repository with
  448. 14:19tons of useful skills, also has a
  449. 14:21handoff skill like that, which I'd
  450. 14:23recommend to you since it does exactly
  451. 14:25that. I'll link it for you in the video
  452. 14:27description. That brings us to the next
  453. 14:29mistake, which is that you really want
  454. 14:30to use AI in a business context, either
  455. 14:32in your own business or as a service.
  456. 14:35You want to understand Claude Code from
  457. 14:37start to finish, along with tools like
  458. 14:39n8n and all the related practices, but
  459. 14:41you keep jumping from YouTube video to
  460. 14:43YouTube video without a real roadmap or
  461. 14:45a structured approach. And you’re
  462. 14:48simply missing the network, meaning the
  463. 14:49exchange with people who are working on
  464. 14:51the same things and from whose mistakes
  465. 14:52and experiences you can learn. But
  466. 14:55don't worry, I have the perfect
  467. 14:56solution for you, and you can drop by
  468. 14:57my community at any time. There, we
  469. 14:59have numerous courses on Claude Code,
  470. 15:01n8n, GDPR, self-hosting, and other
  471. 15:03truly relevant things in a business
  472. 15:05context. And you can exchange ideas
  473. 15:07with me and others here who deal with
  474. 15:08this topic every single day. This way,
  475. 15:10you’ll progress much faster than if
  476. 15:12you have to search for everything
  477. 15:13yourself and repeat the same mistakes
  478. 15:15that someone else might have already
  479. 15:16made. If you're interested, feel free
  480. 15:18to stop by. Link for that is in the
  481. 15:19video description. And now let’s look
  482. 15:21at the next major mistake. Mistake
  483. 15:23number 7, which many people make, is
  484. 15:25not telling Claude Code exactly where
  485. 15:27to look for specific information. For
  486. 15:30example, you open a project folder with
  487. 15:32numerous subfolders and files, and one
  488. 15:34of those subfolders is, say, where you
  489. 15:36keep your quotes, and then you tell
  490. 15:37Claude something like, "Take a look at
  491. 15:39how I write my quotes and how they’re
  492. 15:41structured," but you don't say where it
  493. 15:43can find them. This means Claude now
  494. 15:46has to figure out for itself where this
  495. 15:47quotes folder is, or even if it’s
  496. 15:49named that. It has to read through
  497. 15:51several files that might not even be
  498. 15:52relevant to the context. Anthropic
  499. 15:54calls this "infinite exploration"
  500. 15:56because Claude just starts reading, and
  501. 15:58as we just saw, every file Claude reads
  502. 16:00ends up entirely in its context window.
  503. 16:02This means, in the worst case, it reads
  504. 16:04through random files that have nothing
  505. 16:05to do with what you want to achieve,
  506. 16:07simply because it couldn't precisely
  507. 16:08find the location it was supposed to
  508. 16:09look. So, tell Claude where to look
  509. 16:12whenever possible and definitely use
  510. 16:14that "add" function so you can easily
  511. 16:16select specific files and folders. As
  512. 16:19you can see, many of the points cover
  513. 16:21how we can efficiently handle and
  514. 16:22manage the context that Claude needs to
  515. 16:24process, so we get the best possible
  516. 16:26results and don't unnecessarily clutter
  517. 16:28the model. Let's move on to the next
  518. 16:30mistake, which might actually cost you
  519. 16:32the most of your quota. In Claude Code
  520. 16:34down here on the right, there is an
  521. 16:37effort slider, which goes from low to
  522. 16:40medium, high, extra, max, and even
  523. 16:43ultra code, but the default here is set
  524. 16:46to high. This parameter determines how
  525. 16:48thoroughly Claude works, meaning how
  526. 16:50many files it reads, how many steps it
  527. 16:51takes, and how much it cross-checks in
  528. 16:53the end. In a blog post, Anthropic
  529. 16:56shows an example where the same prompt
  530. 16:57on the high setting generates about
  531. 16:59seven times as many tokens as it does
  532. 17:01on low. And many people just set Claude
  533. 17:04Code to max effort, leave it there
  534. 17:05forever, and then wonder why their
  535. 17:07quota is gone after two hours. Yet,
  536. 17:09Anthropic warns in their own
  537. 17:11documentation that Claude Code tends to
  538. 17:13overthink at max effort and the extra
  539. 17:15cost often yields very little. The rule
  540. 17:17of thumb for what you should use and
  541. 17:19when is also in the Claude
  542. 17:20documentation. For daily work, just
  543. 17:22stay on high. For routine tasks, go to
  544. 17:25medium or even low. And save extra high
  545. 17:28for the really big things. Anthropic
  546. 17:31means tasks where Claude might work for
  547. 17:33more than half an hour at a time. That
  548. 17:37means max and ultra code are totally
  549. 17:38overkill for most things, and ultra
  550. 17:40code isn't even a simple model effort
  551. 17:42level, but a separate feature in Claude
  552. 17:44Code. There, the model runs on extra
  553. 17:48high and, for larger tasks, builds
  554. 17:49entire teams of sub-agents that all
  555. 17:51work together to solve the task you
  556. 17:53give them. That means it's powerful, of
  557. 17:56course, but also very expensive and
  558. 17:57really drains your quota. I would only
  559. 17:59use that for the most complex tasks. So
  560. 18:02just leave it on high, and you'll
  561. 18:04notice you have significantly more of
  562. 18:05your quota left over. And if you ever
  563. 18:08really have a very complex task where
  564. 18:10you notice that even Opus or Claude 3.5
  565. 18:13are having trouble, then bump it up to
  566. 18:15extra or even max. But for most of the
  567. 18:18time, high is absolutely sufficient.
  568. 18:20Let's move on to error number 9, which
  569. 18:22is that you're likely using outdated
  570. 18:24skills that are simply written too
  571. 18:26restrictively for the new models, like
  572. 18:28Opus 3.5 and Sonnet 3.5, and that you
  573. 18:30should also actively maintain your
  574. 18:32Claude file. You already know that the
  575. 18:35Claude MD file should really be as
  576. 18:37short as possible and only contain the
  577. 18:38essential information that Claude
  578. 18:40cannot deduce on its own. Anthropic
  579. 18:42always gives a guideline of under 200
  580. 18:44lines. And if you notice that Claude
  581. 18:46isn't doing certain things, even though
  582. 18:48they are in the Claude MD, then the
  583. 18:50file is simply too long; you should
  584. 18:52shorten it or check again if there are
  585. 18:55any ambiguities in the descriptions you
  586. 18:57included. And with the Claude MD file,
  587. 18:59it's really the case that you don't
  588. 19:01just create it once and never touch it
  589. 19:02again; instead, you should always check
  590. 19:04if you can improve something, or if
  591. 19:05Claude makes mistakes even though it is
  592. 19:07written in the Claude MD. It explicitly
  593. 19:10says here: Treat the Claude MD like
  594. 19:12code, review it when errors occur,
  595. 19:13shorten it regularly, and test changes
  596. 19:15by observing whether Claude's behavior
  597. 19:17actually changes. And if certain things
  598. 19:20are really very important, you can also
  599. 19:22mark them as important in the Claude MD
  600. 19:23. But the best thing you can do is to
  601. 19:26shorten the Claude MD. Because the more
  602. 19:29instructions the language model has to
  603. 19:30process, the higher the probability
  604. 19:32that it will simply overlook or forget
  605. 19:33a specific instruction. And there is a
  606. 19:36great way to customize your Claude MD
  607. 19:38file together with Claude and make it
  608. 19:39better. And that is the slash-do
  609. 19:41command. It's like a diagnostic tool
  610. 19:44that you should really run every now
  611. 19:45and then, and you can simply write
  612. 19:47something like "How can I make my
  613. 19:49Claude MD better?" and now Claude will
  614. 19:51check how the Claude MD might be
  615. 19:52shortened. Can certain information
  616. 19:54perhaps be moved to other files that
  617. 19:56are only read when needed, so that this
  618. 19:58information doesn't end up in the
  619. 20:00context as additional instructions
  620. 20:02every single time? Are there perhaps
  621. 20:04ambiguities that need to be cleared up?
  622. 20:06You can do the same with skills,
  623. 20:08because the Claude documentation states
  624. 20:10that skills written for older models
  625. 20:12are often too prescriptive and can even
  626. 20:14worsen the result. Boris Cherny, who
  627. 20:16built Claude Code at Anthropic, puts it
  628. 20:18even more extremely. He says to delete
  629. 20:21your Claude MD, your skills, and your
  630. 20:23hooks every 6 months and see what the
  631. 20:25model can do without them. Because
  632. 20:27every old instruction was once a patch
  633. 20:28for a mistake that today's model might
  634. 20:30not even make anymore. That doesn't
  635. 20:32mean you shouldn't use skills or hooks,
  636. 20:34but you should simply check if they
  637. 20:36could be written more concisely now, as
  638. 20:38the models have become so good that you
  639. 20:40no longer need to explain everything
  640. 20:42individually. And I see here, for
  641. 20:44example, in Claude Code, that my Claude
  642. 20:46MD has been analyzed, but also other
  643. 20:48issues were found thanks to this slash
  644. 20:50doctor command, using three plugins,
  645. 20:51nine Hickfield skills, and 11
  646. 20:53additional skills, plus the outdated
  647. 20:54Claude MD. That means you should
  648. 20:57definitely do this for your Claude MD.
  649. 20:59Also, go through your current skills
  650. 21:00with Claude Code and see if they can be
  651. 21:02written more concisely or with less
  652. 21:04preamble, as this allows you to
  653. 21:05leverage the strengths of the new
  654. 21:07models. Let's move on to the tenth
  655. 21:09mistake: many people don't know that
  656. 21:11Auto Memory exists. It used to be
  657. 21:13standard to write every insight into
  658. 21:15the Claude MD yourself, and that's
  659. 21:16exactly how those files grow. But now,
  660. 21:19Claude Code has its own memory that
  661. 21:20writes along automatically, and it is
  662. 21:22also turned on by default. These memory
  663. 21:25files are located in your user folder
  664. 21:27under .claude/projects. There is a
  665. 21:30projects folder and inside it a
  666. 21:32subfolder called memory. But you can
  667. 21:34also just ask Claude. Ask it to show
  668. 21:35you what it has remembered about this
  669. 21:38project over time. And alongside the
  670. 21:40Claude MD file, we also see that the
  671. 21:41Auto Memory is loaded into the context
  672. 21:43at the beginning of every session. I am
  673. 21:45here, for example, in a memory folder
  674. 21:47for a project I work on very regularly,
  675. 21:49and here I see many notes that Claude
  676. 21:51has created. And I can now go through
  677. 21:53these notes again and check: Is what it
  678. 21:56remembered still relevant? Is it even
  679. 21:58accurate? And then I can decide: What
  680. 21:59gets removed? What should perhaps be
  681. 22:01explicitly written into the Claude MD?
  682. 22:03What can remain just a memory? And then
  683. 22:05you can simply delete it from the
  684. 22:07Claude MD. So, feel free to take a look
  685. 22:09at what Claude Code has remembered
  686. 22:10about your projects, whether it's all
  687. 22:12accurate, and if there are any
  688. 22:13contradictions between the Claude MD in
  689. 22:15the project and the memories. And
  690. 22:17perhaps you've often wondered why
  691. 22:18Claude keeps doing that one thing over
  692. 22:20and over, even though it's not in the
  693. 22:21Claude MD? Then it could be that it's
  694. 22:24in the memory, and Claude happened to
  695. 22:26remember it by accident at some point.
  696. 22:29Meaning, notice that cloudmd isn't the
  697. 22:31only place to define instructions;
  698. 22:33there’s also the memory that Claude
  699. 22:35Code writes itself while observing your
  700. 22:37work. And if you don't want to look at
  701. 22:39every file individually, you can just
  702. 22:40tell Claude: show me everything you've
  703. 22:42remembered about this project (in
  704. 22:43parentheses: automemory) and compare it
  705. 22:45with the cloudmd for contradictions or
  706. 22:47ambiguities. And Claude has now given
  707. 22:49me a huge analysis here of all the
  708. 22:50memories that exist, which instructions
  709. 22:52contradict each other, where there are
  710. 22:54ambiguities, and so on. And I can now
  711. 22:56fix all of that together with it. You
  712. 22:58can also tell it to write out every
  713. 23:00memory file so you can read through
  714. 23:02them, or you can simply look into the
  715. 23:04files here. This brings us to error
  716. 23:06number 11, which is not giving Claude
  717. 23:08Code the ability to verify its results
  718. 23:10against anything. Yet that is the most
  719. 23:12important tip that Anthropic provides
  720. 23:14here for using Claude Code. The most
  721. 23:16important tip in this guide is
  722. 23:18verification. Give Claude the ability
  723. 23:20to check its own output. When Claude
  724. 23:23has the ability to check its work, the
  725. 23:24quality of the final result is
  726. 23:25significantly improved. Because when
  727. 23:28Claude knows how to check its result,
  728. 23:30it will stay in this agentic loop until
  729. 23:33it actually has the result you want.
  730. 23:35That's why it's best to always provide
  731. 23:37Claude with something it can use to
  732. 23:38compare the result. Some model
  733. 23:40solutions, old outputs, like my old
  734. 23:42newsletters from the very beginning, or
  735. 23:44if you're developing specific programs,
  736. 23:46have tests for every function that
  737. 23:48Claude can check; only when the test is
  738. 23:50successful is the result correct. In
  739. 23:53other words, don't just tell Claude to
  740. 23:54check its result, but tell it how it
  741. 23:56should be checked. And that brings us
  742. 23:57to the last error, and it’s something
  743. 23:59that confuses a lot of people. You
  744. 24:01build a skill, for example in Claude
  745. 24:03Cowork, and notice later that it isn't
  746. 24:05callable in Claude Code, or vice versa.
  747. 24:08That’s not your fault. It’s
  748. 24:10currently just not well-integrated
  749. 24:11between Cowork and Claude Code, because
  750. 24:13Claude Code reads skills from folders
  751. 24:15on your computer, specifically the
  752. 24:17.claude folder, where your skills are
  753. 24:18stored in your directory. For example,
  754. 24:21I have opened the folder where I always
  755. 24:23work, and here is this .claude folder,
  756. 24:26and inside are the skills I use. By the
  757. 24:28way, you have this folder once at the
  758. 24:30project level in such a folder, but
  759. 24:31also at the user level. So, you
  760. 24:33basically have a point cloud folder
  761. 24:35that applies across all projects. That
  762. 24:38means you have global skills that you
  763. 24:40can call up here in every project.
  764. 24:42However, if you're working in Cowork
  765. 24:44alongside Cloud Code, you won't find
  766. 24:46those skills there by default. If you
  767. 24:49click the plus sign here and go to
  768. 24:50Skills, you can see the skills that are
  769. 24:52available in Cowork. The problem with
  770. 24:54Cowork is that it doesn't read the same
  771. 24:57folders as Cloud Code. It only reads
  772. 24:59the skills you've created here in Cloud
  773. 25:01AI, for example in the desktop app or
  774. 25:03in the browser. So, if a skill is to
  775. 25:05run in both Cowork and Cloud Code, you
  776. 25:08have to install it twice. This means
  777. 25:10you have to upload the skill in the app
  778. 25:11if it's not already there. Or you go to
  779. 25:14Cloud Code and insert the skill file
  780. 25:16you either uploaded to Cloud Cowork or
  781. 25:18created with Cloud, and tell Cloud Code
  782. 25:20to install it in this project. And then
  783. 25:23you can call up the skills here using a
  784. 25:24slash command. I would generally
  785. 25:27recommend working in Cloud Code anyway,
  786. 25:29but if you like using both, you have to
  787. 25:31ensure the skills you need are
  788. 25:33available in both. That means you have
  789. 25:36to install them twice. Unfortunately,
  790. 25:37that’s not very well solved at the
  791. 25:38moment. By the way, hooks only work in
  792. 25:41Cloud Code, not in Cowork, but
  793. 25:43something like a Cloud MDEI works in
  794. 25:45both Cloud Code and Cowork. That is
  795. 25:48always read as soon as you work within
  796. 25:50a folder with a Cloud MDEI file. Those
  797. 25:52were the 12 points I wanted to go over
  798. 25:53with you. I hope you were able to take
  799. 25:55a few things away and that I was able
  800. 25:57to clear up a few questions for you. I
  801. 25:59have written down the essence of all
  802. 26:00the points for you here again. And if I
  803. 26:03can highlight three things, it is
  804. 26:04definitely the verification of results
  805. 26:06that Clot can take over itself, and
  806. 26:08that you explain the reason for your
  807. 26:10plan and your goal to Clot when
  808. 26:12prompting, instead of telling it
  809. 26:13exactly what to do. Secondly, maintain
  810. 26:17Cloud's context window, keep it as
  811. 26:18clear as possible, start new sessions
  812. 26:20for new tasks, don't jump back and
  813. 26:21forth between models, and use commands
  814. 26:23like the slash compact command if
  815. 26:24you've been working on a task for a
  816. 26:26very long time. And thirdly, keep your
  817. 26:29setup lean in general; keep your
  818. 26:31Cloudmds as short as possible with
  819. 26:32clear instructions, and check your
  820. 26:34skills again to see if they might be
  821. 26:36written too restrictively for the new
  822. 26:38models, preventing you from unlocking
  823. 26:40their full potential. That's it for
  824. 26:43this video. Feel free to write in the
  825. 26:44comments what other tips you have or
  826. 26:46mistakes one should perhaps avoid. If
  827. 26:48you want to see more content on Cloud
  828. 26:50Code, definitely leave a like so I know
  829. 26:52, and feel free to subscribe as it's
  830. 26:54free for you and ensures you won't miss
  831. 26:55any future content. Thank you very much
  832. 26:58for watching, and I'll see you in the
  833. 26:59next video. See you then.

About this transcript

This page contains the full transcript of So nutzt du Claude Code nach Anthropics neuen Regeln by Julian Ivanov | KI-Automatisierung, generated from the public captions YouTube serves with the video. The transcript has 5,623 words across 833 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.