So nutzt du Claude Code nach Anthropics neuen Regeln — Transcript
Full transcript
- 0:00This summer, Anthropic deleted about 80
- 0:02%of Claude's system prompt—the
- 0:04instructions that Claude automatically
- 0:06receives with every single message—
- 0:08and according to Anthropic, the model
- 0:10has actually become better because of
- 0:12it. At the same time, the documentation
- 0:15was rewritten in quite a few places,
- 0:16because with Opus 3.5 and Sonnet 3.5,
- 0:18the models behave differently than they
- 0:20did just half a year ago. This means
- 0:22that some of the habits you've picked
- 0:24up might not be entirely optimal as of
- 0:26today and could actually be slowing
- 0:28Claude down. That's why today I'm going
- 0:31to go through 12 mistakes that many
- 0:32people still make when working with
- 0:34Claude, and for each one, I'll show you
- 0:36what you should do instead. Claude
- 0:37Anthropic. Let's start from the very
- 0:39beginning with the pure prompting of
- 0:41these new models. The models have
- 0:44become so good by now that we also need
- 0:46to adjust the way we prompt to really
- 0:48get the quality these models can
- 0:49deliver. This applies not just to
- 0:52Claude, but also here in the normal
- 0:54Claude chat window or in Projects.
- 0:56Let's look at a concrete example.
- 0:58Assume I want to write a newsletter for
- 1:00my community and I want Claude to do it
- 1:02for me. Many people write the prompt in
- 1:05roughly this style: "You are an
- 1:06excellent newsletter author with years
- 1:08of experience. Write me this week's
- 1:11newsletter on the topic of new best
- 1:12practices for Claude. Write in my style
- 1:15, keep it casual, and not too long."
- 1:16Even the top models will still have
- 1:18trouble with a prompt like that and
- 1:20won't achieve the result you might be
- 1:22imagining. Just look at everything the
- 1:24model has to guess with this prompt.
- 1:26First, who the newsletter is for and
- 1:27what the readers should do with it.
- 1:29Then, what does "my style" even mean,
- 1:31and what does "not too long" mean? We
- 1:34also aren't giving it any reference
- 1:36material here that would give Claude a
- 1:37clue about how we like to write
- 1:39newsletters. Additionally, it has been
- 1:41discovered that phrases like "you are
- 1:43an excellent newsletter author with
- 1:45years of experience" do nothing to help
- 1:47the model provide better output. Things
- 1:50like that are actually ignored by the
- 1:51model, and the only thing it takes from
- 1:53that role is simply the style. So, if I
- 1:56say "you are a Romantic-era author,
- 1:57write the following text," the model is
- 1:59already trying its best to give you the
- 2:02best possible output if you just say
- 2:03you need a newsletter. It doesn’t
- 2:05help anymore if you say, yes, this is
- 2:07extremely important or I need your full
- 2:08concentration or something like that.
- 2:11Anthropic has created a prompt library
- 2:13in their documentation with a section
- 2:14on what actually makes a good prompt.
- 2:17And you can really break the whole
- 2:19thing down into three essential points.
- 2:20And the first is that you should never
- 2:22just give the pure instruction or
- 2:24request to the model, but always also
- 2:26the reason for it and your goal. The
- 2:29documentation says the model works
- 2:31better if it knows the intention behind
- 2:33the task, because then it can connect
- 2:34the task to what else it knows, instead
- 2:36of guessing your intention. And the
- 2:40template there effectively says: I am
- 2:42currently working on this task for this
- 2:44target audience. They need the
- 2:46following output, and with this
- 2:48information, please carry out the
- 2:49following task. Second, it is also
- 2:52usually better to show Claude something
- 2:53instead of describing it. If I write
- 2:55something to Claude like I write in a
- 2:57casual style and not too long, then it
- 2:59is incredibly difficult for Claude to
- 3:01replicate my style if I haven't defined
- 3:03it anywhere in a file. When you work in
- 3:05Claude Code, you can use an@to
- 3:07reference specific files that Claude
- 3:08Code then reads directly along with the
- 3:10prompt. And here, for example, it would
- 3:13make a lot of sense to simply reference
- 3:15past newsletters that I have written
- 3:16myself. Claude can then read through
- 3:18them, and then this newsletter will
- 3:20have the new writing style. That will
- 3:21work significantly better than if you
- 3:23try to describe your way of writing to
- 3:25it somehow. And the third point, which
- 3:27is also extremely important, is what
- 3:28does finished even mean and how can
- 3:30Claude check the result itself?
- 3:31According to the documentation, you
- 3:33describe the result with a measurable
- 3:35goal so that it is clear when Claude
- 3:37can stop. So, if I now write something
- 3:39like the newsletter shouldn't be too
- 3:40long, that’s not a real number. But
- 3:42if I say 500 words, three to four
- 3:44paragraphs, and a small tip at the end,
- 3:46then that is much more measurable, and
- 3:48Claude can do much more with it, and
- 3:50then you give Claude the opportunity to
- 3:52check its result itself. This is the
- 3:54first essential best practice in the
- 3:56entire best practice documentation here
- 3:58from Claude Code, because the models
- 3:59have become much better at working on a
- 4:01task for a long time. This means Claude
- 4:03Code can stay in this agentic loop
- 4:05until the result is exactly how you
- 4:07want it to be. If you are just writing
- 4:09a longer text now, you can also say
- 4:11something like: "Look at the past
- 4:13newsletters now." e.g., the ones we
- 4:15wrote, at the end, and check if the new
- 4:17newsletter you generated matches the
- 4:19old ones—in tone, length, content,
- 4:22and so on. This means it’s not enough
- 4:24to just tell it: "Yes, check yourself
- 4:25at the end, check your result," but you
- 4:27must also tell it what to check against
- 4:29—what are the evaluation criteria?
- 4:30And last but not least, it can be very
- 4:32useful to tell the model what you do
- 4:34not expect in the result or what you do
- 4:36not want it to do. Because especially
- 4:38Opus 3.5 and Sonnet 3.5 often tend to
- 4:40come up with additional tasks that you
- 4:42didn't assign at the beginning, and
- 4:44then they do something for you, and you
- 4:45think, why did you just do that? I
- 4:48didn't even ask for that, and how did
- 4:50you even figure out you could do that,
- 4:52and so on. It even says literally in
- 4:54the documentation here for Opus 3.5,
- 4:56for example, because the model wants to
- 4:58take everything off your hands as much
- 4:59as possible, but there are certain
- 5:01things you explicitly don't want. And
- 5:03if you know that, if you know what you
- 5:05don't want in the result, then you
- 5:07should clearly write that in the prompt
- 5:08, otherwise the model will just do it
- 5:10because it thinks it would be useful
- 5:11for you. That means a much better
- 5:13prompt than the one at the beginning
- 5:14would be something like this. I write a
- 5:16newsletter for my community every week.
- 5:18They are primarily self-employed
- 5:19entrepreneurs and AI enthusiasts. Many
- 5:21are already working with Claude. Every
- 5:23week, they should take away a few best
- 5:24practices and news items that they can
- 5:26implement directly in their work with
- 5:27AI. With that, we immediately have the
- 5:29background, the target group, and what
- 5:31we want to achieve. Then I give it the
- 5:33reference. The topic this week is the
- 5:35new best practices for Claude. My notes
- 5:38on this are in this file. You can't
- 5:41reference that directly here in the
- 5:42Claude chat. That works in Claude for
- 5:44developers. And for the style and
- 5:46structure, take the last three issues
- 5:47—and then I give it the folder with
- 5:49my last newsletters, for example. At
- 5:51the end, I also say what "finished"
- 5:52means to me: namely, when the topic is
- 5:54explained in three to four paragraphs
- 5:56and there is one thing at the end that
- 5:58you can try out this week. Compare the
- 6:00tone and length at the end with the
- 6:02three issues, and I only want the
- 6:03newsletter text, no subject line or any
- 6:05other things. So, at the beginning, we
- 6:07have the background and the goal. We
- 6:09have references that Claude can use,
- 6:11and we define what our final result
- 6:12should look like and how it can verify
- 6:14it. Believe me, if you follow these
- 6:17three things when working with agentic
- 6:18systems like Claude, Hermes, Codex, and
- 6:20so on, you will get significantly
- 6:22better results. And since this is a
- 6:25task I do every week, it would be very
- 6:27smart to turn it into a skill once
- 6:29Claude has done it with me and we are
- 6:31happy with the result. That is then
- 6:34essentially a set of instructions that
- 6:35Claude can use every time I want to
- 6:36write a newsletter. I can then call up
- 6:38this skill every week when I write a
- 6:40newsletter, and Claude immediately
- 6:42knows the drill and what it needs to do
- 6:44, and I can additionally provide it
- 6:45with references for the latest topics
- 6:47and so on. Let's move on to the second
- 6:49mistake, which is that many people
- 6:51still switch models in the same chat
- 6:53here and there while chatting with
- 6:55Claude to save a few tokens. Claude
- 6:57sends the entire previous chat history
- 7:00with every message. But so that this
- 7:02doesn't fully impact your quota every
- 7:03time, there is a so-called cache. This
- 7:05means that what has already been
- 7:07processed only counts as a fraction the
- 7:09next time. And this cache is tied to
- 7:11the specific model you are currently
- 7:12using. However, if you switch models in
- 7:15the middle of a chat, the entire chat
- 7:17history is re-uploaded to the model,
- 7:19and all those tokens are consumed once
- 7:21again. And that naturally eats into
- 7:24your quota, because all those tokens
- 7:25have to be fully processed again. And
- 7:27that is the reason why some people hit
- 7:29their session limit after just two
- 7:30hours, even though they haven't done as
- 7:32much as they claim. This means you
- 7:34should avoid switching back and forth
- 7:35between models too much. It is best to
- 7:38work with one model on one task, finish
- 7:40it, and if you realize you want to do
- 7:42something else, start a new task, then
- 7:44start a new chat with that specific
- 7:46model, be it Opus, Sonnet, or Haiku.
- 7:49Let's come to the third mistake, which
- 7:51is telling Claude too often what it
- 7:53should not do. That was a big point
- 7:55that Anthropic also shortened in the
- 7:57Claude documentation for Claude,
- 7:59because they had specific instructions
- 8:01in many places about what the model
- 8:03should not do. And the problem is, if
- 8:05you keep telling it what not to do,
- 8:07you'll confuse the model in the long
- 8:09run, because the more prohibitions you
- 8:11define, the more it has to weigh each
- 8:12individual prohibition against the rest
- 8:14of the instructions. And Anthropic has
- 8:16realized that it makes more sense,
- 8:18instead of giving the model a lot of
- 8:19prohibitions, to rather say exactly
- 8:21what it should do concretely. That
- 8:23means, instead of saying don't be too
- 8:24long-winded, it is better to write,
- 8:26answer in no more than five sentences.
- 8:28Or instead of writing no bullet points,
- 8:29I prefer to write, write in full
- 8:31sentences. And if you have very
- 8:32specific things that it really
- 8:34shouldn't do, then that's no problem
- 8:35either, you can of course tell it that.
- 8:37But the point here is really just that
- 8:38you should rather tell Claude what it
- 8:40should do, instead of what it should
- 8:42not do, so that you simply avoid such a
- 8:43list of prohibitions, which at some
- 8:45point will contradict each other and
- 8:47just confuse Claude, because it then
- 8:48has to deduce what it should do instead
- 8:50. And with that, we come to the fourth
- 8:52mistake. If we want Claude to really
- 8:54always do something specific, most
- 8:56people just write another rule in the
- 8:58Claude.md file along the lines of
- 9:00please do it this way every time. But
- 9:03the problem with that is that this
- 9:04Claude.md is ultimately just additional
- 9:06context for Claude. Just like the
- 9:08system prompt, which is sort of already
- 9:09built into Claude's code by Anthropic.
- 9:11That is also always loaded into the
- 9:12context. Just like the tools and skills
- 9:14that Claude can use. Just like the
- 9:16entire chat history and also the files
- 9:18that Claude has read. All this
- 9:20information is loaded into Claude's
- 9:22context, and the longer you work on the
- 9:24task and the more the context window
- 9:26fills up, the more often it can happen
- 9:28that instructions are simply not
- 9:29followed because they get lost under
- 9:31all that context. This means the
- 9:33Claude.md file is not a set of rules
- 9:34that always applies and that Claude
- 9:36always follows. If I write in my
- 9:38Claude.md, for example, always write
- 9:39German texts with umlauts, then that
- 9:41will happen most of the time, but there
- 9:43will always be cases where it forgets.
- 9:45Especially when I've been working on
- 9:46the task for a longer time. But if you
- 9:47really want to ensure that a certain
- 9:49action is always carried out, then you
- 9:51should definitely use hooks. Hooks
- 9:53automatically execute scripts at
- 9:55certain points in Claude's workflow.
- 9:57Unlike the instructions in the
- 9:59Claude.md, which are merely
- 10:01recommendations, hooks are
- 10:02deterministic and guarantee the
- 10:04execution of the action. So, if I want
- 10:07to ensure that Claude......for example,
- 10:09when creating files and including
- 10:11German text, actually writes umlauts
- 10:13correctly, I create a hook for that,
- 10:14which then deterministically checks
- 10:16every time that it happens. And if not,
- 10:19then Claude is notified and has to
- 10:21rewrite the text. If I want to ensure,
- 10:24for instance, that Claude doesn't look
- 10:25at certain files because they contain
- 10:27credentials or other data that Claude
- 10:29shouldn't read, I can set up a hook for
- 10:31that as well. Because as soon as Claude
- 10:34tries to read the file, this hook is
- 10:35activated and blocks the action. And if
- 10:38you are wondering now how to create
- 10:39such a hook, you just have to tell
- 10:40Claude Code. Just tell it: "Hey, I
- 10:42would like you to create a hook that
- 10:44ensures that this and that always
- 10:45happens in such a way." This really
- 10:48ensures that it happens, and you don't
- 10:50have to rely on Claude having
- 10:51remembered it from the instructions
- 10:53within the whole context it has to
- 10:54process in parallel. And let's stick
- 10:57with the Claude.md file, because many
- 10:58people still make mistakes here. And
- 11:00the fifth mistake is changing the
- 11:01Claude.md file in the middle of a
- 11:03session. If you are working on a task
- 11:05with Claude Code and you notice that
- 11:07Claude isn't doing something right—it
- 11:09keeps getting it wrong—and you want
- 11:11to add that to the Claude.md, you add
- 11:13it, but then continue in the chat, the
- 11:15new version of this Claude.md file will
- 11:17not be loaded. This file is only ever
- 11:20loaded once at the beginning of the
- 11:22session. That means any changes made
- 11:23afterward are not taken into account in
- 11:25this session. This leads to some people
- 11:27complaining that Claude isn't following
- 11:28the instructions despite them being in
- 11:30the file. But that is because you
- 11:31simply have to start a new session so
- 11:33that the new Claude.md file is taken
- 11:35into account. This brings us to the
- 11:37sixth mistake, which is that people
- 11:38still underestimate how important it is
- 11:40to keep the context window of these
- 11:42language models as empty as possible.
- 11:45The more information the language model
- 11:46has to process at the same time and the
- 11:48more the context window is filled, the
- 11:50dumber it becomes and the worse it gets
- 11:52at retrieving specific information. It
- 11:54simply loses track. Anthropic measures
- 11:56this with a benchmark called Needle In
- 11:58A Haystack. There, various pieces of
- 12:00information are hidden in a huge text,
- 12:02and then it is tested how many of them
- 12:03the model can find again. And by now,
- 12:05the Claude models have a context window
- 12:07of one million tokens. That means it
- 12:09can process 1 million tokens, or about
- 12:12750,000 words, at the same time. But
- 12:14that doesn't mean the model can
- 12:16actually retrieve every piece of
- 12:17information, because at the end of the
- 12:19day, it's all based on probabilities.
- 12:21And it's been shown that when the model
- 12:24processes only about 256,000 tokens, it
- 12:26finds 93%of the requested information,
- 12:29but as soon as it processes 1 million
- 12:31tokens, that drops to 76%. This means
- 12:36in a packed chat, when the context
- 12:37window is completely full, it can
- 12:39happen that on average every fourth
- 12:41piece of information is lost or Claude
- 12:43simply loses track of it, and that
- 12:45includes your rules and instructions.
- 12:48So it's really worth regularly starting
- 12:50a new chat, clearing the context, or
- 12:53using the slash-compact command to
- 12:55summarize the conversation history and
- 12:57the entire context to continue working.
- 13:00Because just take a look at this
- 13:01context window simulation from Claude
- 13:03Code here. Here the documentation shows
- 13:05how much context Claude Code actually
- 13:07processes with every session we start.
- 13:09And as soon as I start the session now,
- 13:11the following things are loaded into
- 13:13the language model's context window.
- 13:15System prompt, auto-memory, environment
- 13:17info, MCP tools, skills, the Claude.md,
- 13:19and all that before I've even written
- 13:21anything. And when I then write a
- 13:23message, like "fix this bug," my prompt
- 13:26is added here, but also every file that
- 13:29Claude reads through. That means just
- 13:31after this one message, I've already
- 13:33used up 21,000 tokens, and people
- 13:34underestimate how fast that happens. By
- 13:37the way, you can type/context in the
- 13:38chat with Claude Code at any time to
- 13:40see how the context window is currently
- 13:42filled, which often shows you what you
- 13:44might be able to optimize. So, really
- 13:47get into the habit of starting a new
- 13:49chat or using/clear when you're done
- 13:51with a task, or using something like/
- 13:53compact. Then you keep a summary of
- 13:55what was done before. You can also
- 13:57define what you want to have summarized
- 13:59. Or what you can also do—many people
- 14:01do this—is have a handoff written for
- 14:03the current session. That just means a
- 14:06summary in a file that Claude can then
- 14:08read when you start a new session. This
- 14:10is also a great skill to have, right?
- 14:12Then next time you can just write/
- 14:14handoff, for example, and Claude will
- 14:15create it automatically. Mad Pocock,
- 14:18who has a really great repository with
- 14:19tons of useful skills, also has a
- 14:21handoff skill like that, which I'd
- 14:23recommend to you since it does exactly
- 14:25that. I'll link it for you in the video
- 14:27description. That brings us to the next
- 14:29mistake, which is that you really want
- 14:30to use AI in a business context, either
- 14:32in your own business or as a service.
- 14:35You want to understand Claude Code from
- 14:37start to finish, along with tools like
- 14:39n8n and all the related practices, but
- 14:41you keep jumping from YouTube video to
- 14:43YouTube video without a real roadmap or
- 14:45a structured approach. And you’re
- 14:48simply missing the network, meaning the
- 14:49exchange with people who are working on
- 14:51the same things and from whose mistakes
- 14:52and experiences you can learn. But
- 14:55don't worry, I have the perfect
- 14:56solution for you, and you can drop by
- 14:57my community at any time. There, we
- 14:59have numerous courses on Claude Code,
- 15:01n8n, GDPR, self-hosting, and other
- 15:03truly relevant things in a business
- 15:05context. And you can exchange ideas
- 15:07with me and others here who deal with
- 15:08this topic every single day. This way,
- 15:10you’ll progress much faster than if
- 15:12you have to search for everything
- 15:13yourself and repeat the same mistakes
- 15:15that someone else might have already
- 15:16made. If you're interested, feel free
- 15:18to stop by. Link for that is in the
- 15:19video description. And now let’s look
- 15:21at the next major mistake. Mistake
- 15:23number 7, which many people make, is
- 15:25not telling Claude Code exactly where
- 15:27to look for specific information. For
- 15:30example, you open a project folder with
- 15:32numerous subfolders and files, and one
- 15:34of those subfolders is, say, where you
- 15:36keep your quotes, and then you tell
- 15:37Claude something like, "Take a look at
- 15:39how I write my quotes and how they’re
- 15:41structured," but you don't say where it
- 15:43can find them. This means Claude now
- 15:46has to figure out for itself where this
- 15:47quotes folder is, or even if it’s
- 15:49named that. It has to read through
- 15:51several files that might not even be
- 15:52relevant to the context. Anthropic
- 15:54calls this "infinite exploration"
- 15:56because Claude just starts reading, and
- 15:58as we just saw, every file Claude reads
- 16:00ends up entirely in its context window.
- 16:02This means, in the worst case, it reads
- 16:04through random files that have nothing
- 16:05to do with what you want to achieve,
- 16:07simply because it couldn't precisely
- 16:08find the location it was supposed to
- 16:09look. So, tell Claude where to look
- 16:12whenever possible and definitely use
- 16:14that "add" function so you can easily
- 16:16select specific files and folders. As
- 16:19you can see, many of the points cover
- 16:21how we can efficiently handle and
- 16:22manage the context that Claude needs to
- 16:24process, so we get the best possible
- 16:26results and don't unnecessarily clutter
- 16:28the model. Let's move on to the next
- 16:30mistake, which might actually cost you
- 16:32the most of your quota. In Claude Code
- 16:34down here on the right, there is an
- 16:37effort slider, which goes from low to
- 16:40medium, high, extra, max, and even
- 16:43ultra code, but the default here is set
- 16:46to high. This parameter determines how
- 16:48thoroughly Claude works, meaning how
- 16:50many files it reads, how many steps it
- 16:51takes, and how much it cross-checks in
- 16:53the end. In a blog post, Anthropic
- 16:56shows an example where the same prompt
- 16:57on the high setting generates about
- 16:59seven times as many tokens as it does
- 17:01on low. And many people just set Claude
- 17:04Code to max effort, leave it there
- 17:05forever, and then wonder why their
- 17:07quota is gone after two hours. Yet,
- 17:09Anthropic warns in their own
- 17:11documentation that Claude Code tends to
- 17:13overthink at max effort and the extra
- 17:15cost often yields very little. The rule
- 17:17of thumb for what you should use and
- 17:19when is also in the Claude
- 17:20documentation. For daily work, just
- 17:22stay on high. For routine tasks, go to
- 17:25medium or even low. And save extra high
- 17:28for the really big things. Anthropic
- 17:31means tasks where Claude might work for
- 17:33more than half an hour at a time. That
- 17:37means max and ultra code are totally
- 17:38overkill for most things, and ultra
- 17:40code isn't even a simple model effort
- 17:42level, but a separate feature in Claude
- 17:44Code. There, the model runs on extra
- 17:48high and, for larger tasks, builds
- 17:49entire teams of sub-agents that all
- 17:51work together to solve the task you
- 17:53give them. That means it's powerful, of
- 17:56course, but also very expensive and
- 17:57really drains your quota. I would only
- 17:59use that for the most complex tasks. So
- 18:02just leave it on high, and you'll
- 18:04notice you have significantly more of
- 18:05your quota left over. And if you ever
- 18:08really have a very complex task where
- 18:10you notice that even Opus or Claude 3.5
- 18:13are having trouble, then bump it up to
- 18:15extra or even max. But for most of the
- 18:18time, high is absolutely sufficient.
- 18:20Let's move on to error number 9, which
- 18:22is that you're likely using outdated
- 18:24skills that are simply written too
- 18:26restrictively for the new models, like
- 18:28Opus 3.5 and Sonnet 3.5, and that you
- 18:30should also actively maintain your
- 18:32Claude file. You already know that the
- 18:35Claude MD file should really be as
- 18:37short as possible and only contain the
- 18:38essential information that Claude
- 18:40cannot deduce on its own. Anthropic
- 18:42always gives a guideline of under 200
- 18:44lines. And if you notice that Claude
- 18:46isn't doing certain things, even though
- 18:48they are in the Claude MD, then the
- 18:50file is simply too long; you should
- 18:52shorten it or check again if there are
- 18:55any ambiguities in the descriptions you
- 18:57included. And with the Claude MD file,
- 18:59it's really the case that you don't
- 19:01just create it once and never touch it
- 19:02again; instead, you should always check
- 19:04if you can improve something, or if
- 19:05Claude makes mistakes even though it is
- 19:07written in the Claude MD. It explicitly
- 19:10says here: Treat the Claude MD like
- 19:12code, review it when errors occur,
- 19:13shorten it regularly, and test changes
- 19:15by observing whether Claude's behavior
- 19:17actually changes. And if certain things
- 19:20are really very important, you can also
- 19:22mark them as important in the Claude MD
- 19:23. But the best thing you can do is to
- 19:26shorten the Claude MD. Because the more
- 19:29instructions the language model has to
- 19:30process, the higher the probability
- 19:32that it will simply overlook or forget
- 19:33a specific instruction. And there is a
- 19:36great way to customize your Claude MD
- 19:38file together with Claude and make it
- 19:39better. And that is the slash-do
- 19:41command. It's like a diagnostic tool
- 19:44that you should really run every now
- 19:45and then, and you can simply write
- 19:47something like "How can I make my
- 19:49Claude MD better?" and now Claude will
- 19:51check how the Claude MD might be
- 19:52shortened. Can certain information
- 19:54perhaps be moved to other files that
- 19:56are only read when needed, so that this
- 19:58information doesn't end up in the
- 20:00context as additional instructions
- 20:02every single time? Are there perhaps
- 20:04ambiguities that need to be cleared up?
- 20:06You can do the same with skills,
- 20:08because the Claude documentation states
- 20:10that skills written for older models
- 20:12are often too prescriptive and can even
- 20:14worsen the result. Boris Cherny, who
- 20:16built Claude Code at Anthropic, puts it
- 20:18even more extremely. He says to delete
- 20:21your Claude MD, your skills, and your
- 20:23hooks every 6 months and see what the
- 20:25model can do without them. Because
- 20:27every old instruction was once a patch
- 20:28for a mistake that today's model might
- 20:30not even make anymore. That doesn't
- 20:32mean you shouldn't use skills or hooks,
- 20:34but you should simply check if they
- 20:36could be written more concisely now, as
- 20:38the models have become so good that you
- 20:40no longer need to explain everything
- 20:42individually. And I see here, for
- 20:44example, in Claude Code, that my Claude
- 20:46MD has been analyzed, but also other
- 20:48issues were found thanks to this slash
- 20:50doctor command, using three plugins,
- 20:51nine Hickfield skills, and 11
- 20:53additional skills, plus the outdated
- 20:54Claude MD. That means you should
- 20:57definitely do this for your Claude MD.
- 20:59Also, go through your current skills
- 21:00with Claude Code and see if they can be
- 21:02written more concisely or with less
- 21:04preamble, as this allows you to
- 21:05leverage the strengths of the new
- 21:07models. Let's move on to the tenth
- 21:09mistake: many people don't know that
- 21:11Auto Memory exists. It used to be
- 21:13standard to write every insight into
- 21:15the Claude MD yourself, and that's
- 21:16exactly how those files grow. But now,
- 21:19Claude Code has its own memory that
- 21:20writes along automatically, and it is
- 21:22also turned on by default. These memory
- 21:25files are located in your user folder
- 21:27under .claude/projects. There is a
- 21:30projects folder and inside it a
- 21:32subfolder called memory. But you can
- 21:34also just ask Claude. Ask it to show
- 21:35you what it has remembered about this
- 21:38project over time. And alongside the
- 21:40Claude MD file, we also see that the
- 21:41Auto Memory is loaded into the context
- 21:43at the beginning of every session. I am
- 21:45here, for example, in a memory folder
- 21:47for a project I work on very regularly,
- 21:49and here I see many notes that Claude
- 21:51has created. And I can now go through
- 21:53these notes again and check: Is what it
- 21:56remembered still relevant? Is it even
- 21:58accurate? And then I can decide: What
- 21:59gets removed? What should perhaps be
- 22:01explicitly written into the Claude MD?
- 22:03What can remain just a memory? And then
- 22:05you can simply delete it from the
- 22:07Claude MD. So, feel free to take a look
- 22:09at what Claude Code has remembered
- 22:10about your projects, whether it's all
- 22:12accurate, and if there are any
- 22:13contradictions between the Claude MD in
- 22:15the project and the memories. And
- 22:17perhaps you've often wondered why
- 22:18Claude keeps doing that one thing over
- 22:20and over, even though it's not in the
- 22:21Claude MD? Then it could be that it's
- 22:24in the memory, and Claude happened to
- 22:26remember it by accident at some point.
- 22:29Meaning, notice that cloudmd isn't the
- 22:31only place to define instructions;
- 22:33there’s also the memory that Claude
- 22:35Code writes itself while observing your
- 22:37work. And if you don't want to look at
- 22:39every file individually, you can just
- 22:40tell Claude: show me everything you've
- 22:42remembered about this project (in
- 22:43parentheses: automemory) and compare it
- 22:45with the cloudmd for contradictions or
- 22:47ambiguities. And Claude has now given
- 22:49me a huge analysis here of all the
- 22:50memories that exist, which instructions
- 22:52contradict each other, where there are
- 22:54ambiguities, and so on. And I can now
- 22:56fix all of that together with it. You
- 22:58can also tell it to write out every
- 23:00memory file so you can read through
- 23:02them, or you can simply look into the
- 23:04files here. This brings us to error
- 23:06number 11, which is not giving Claude
- 23:08Code the ability to verify its results
- 23:10against anything. Yet that is the most
- 23:12important tip that Anthropic provides
- 23:14here for using Claude Code. The most
- 23:16important tip in this guide is
- 23:18verification. Give Claude the ability
- 23:20to check its own output. When Claude
- 23:23has the ability to check its work, the
- 23:24quality of the final result is
- 23:25significantly improved. Because when
- 23:28Claude knows how to check its result,
- 23:30it will stay in this agentic loop until
- 23:33it actually has the result you want.
- 23:35That's why it's best to always provide
- 23:37Claude with something it can use to
- 23:38compare the result. Some model
- 23:40solutions, old outputs, like my old
- 23:42newsletters from the very beginning, or
- 23:44if you're developing specific programs,
- 23:46have tests for every function that
- 23:48Claude can check; only when the test is
- 23:50successful is the result correct. In
- 23:53other words, don't just tell Claude to
- 23:54check its result, but tell it how it
- 23:56should be checked. And that brings us
- 23:57to the last error, and it’s something
- 23:59that confuses a lot of people. You
- 24:01build a skill, for example in Claude
- 24:03Cowork, and notice later that it isn't
- 24:05callable in Claude Code, or vice versa.
- 24:08That’s not your fault. It’s
- 24:10currently just not well-integrated
- 24:11between Cowork and Claude Code, because
- 24:13Claude Code reads skills from folders
- 24:15on your computer, specifically the
- 24:17.claude folder, where your skills are
- 24:18stored in your directory. For example,
- 24:21I have opened the folder where I always
- 24:23work, and here is this .claude folder,
- 24:26and inside are the skills I use. By the
- 24:28way, you have this folder once at the
- 24:30project level in such a folder, but
- 24:31also at the user level. So, you
- 24:33basically have a point cloud folder
- 24:35that applies across all projects. That
- 24:38means you have global skills that you
- 24:40can call up here in every project.
- 24:42However, if you're working in Cowork
- 24:44alongside Cloud Code, you won't find
- 24:46those skills there by default. If you
- 24:49click the plus sign here and go to
- 24:50Skills, you can see the skills that are
- 24:52available in Cowork. The problem with
- 24:54Cowork is that it doesn't read the same
- 24:57folders as Cloud Code. It only reads
- 24:59the skills you've created here in Cloud
- 25:01AI, for example in the desktop app or
- 25:03in the browser. So, if a skill is to
- 25:05run in both Cowork and Cloud Code, you
- 25:08have to install it twice. This means
- 25:10you have to upload the skill in the app
- 25:11if it's not already there. Or you go to
- 25:14Cloud Code and insert the skill file
- 25:16you either uploaded to Cloud Cowork or
- 25:18created with Cloud, and tell Cloud Code
- 25:20to install it in this project. And then
- 25:23you can call up the skills here using a
- 25:24slash command. I would generally
- 25:27recommend working in Cloud Code anyway,
- 25:29but if you like using both, you have to
- 25:31ensure the skills you need are
- 25:33available in both. That means you have
- 25:36to install them twice. Unfortunately,
- 25:37that’s not very well solved at the
- 25:38moment. By the way, hooks only work in
- 25:41Cloud Code, not in Cowork, but
- 25:43something like a Cloud MDEI works in
- 25:45both Cloud Code and Cowork. That is
- 25:48always read as soon as you work within
- 25:50a folder with a Cloud MDEI file. Those
- 25:52were the 12 points I wanted to go over
- 25:53with you. I hope you were able to take
- 25:55a few things away and that I was able
- 25:57to clear up a few questions for you. I
- 25:59have written down the essence of all
- 26:00the points for you here again. And if I
- 26:03can highlight three things, it is
- 26:04definitely the verification of results
- 26:06that Clot can take over itself, and
- 26:08that you explain the reason for your
- 26:10plan and your goal to Clot when
- 26:12prompting, instead of telling it
- 26:13exactly what to do. Secondly, maintain
- 26:17Cloud's context window, keep it as
- 26:18clear as possible, start new sessions
- 26:20for new tasks, don't jump back and
- 26:21forth between models, and use commands
- 26:23like the slash compact command if
- 26:24you've been working on a task for a
- 26:26very long time. And thirdly, keep your
- 26:29setup lean in general; keep your
- 26:31Cloudmds as short as possible with
- 26:32clear instructions, and check your
- 26:34skills again to see if they might be
- 26:36written too restrictively for the new
- 26:38models, preventing you from unlocking
- 26:40their full potential. That's it for
- 26:43this video. Feel free to write in the
- 26:44comments what other tips you have or
- 26:46mistakes one should perhaps avoid. If
- 26:48you want to see more content on Cloud
- 26:50Code, definitely leave a like so I know
- 26:52, and feel free to subscribe as it's
- 26:54free for you and ensures you won't miss
- 26:55any future content. Thank you very much
- 26:58for watching, and I'll see you in the
- 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.