Transcript
0:00 Anthropic just released a new article on how to actually use effort levels. And while many people assume the higher the better and simply never use low or always just put it on high, it's actually much more nuanced than that. And choosing the right effort level can make a big impact on output quality, token consumption, and speed. So, in this video, I'll explain what effort actually does based on Anthropic's own research, show you exactly which level to use for which use case, and show you why you probably need to use low a lot more, and give away a free mod that helps you apply the best effort level to your prompts automatically. Now, if you don't know me yet, I'm Ben. I'm a three-time founder and now run an AI agency and one of the largest AI communities for professionals and business owners where I teach how to use AI to actually automate work or build an AI business. So, what actually are these effort levels? We of course have low, medium, high, extra high, and max. Now, many people think it's about how smart Claude is, but how smart depends much more on the model you pick like Opus, Sonnet, or Haiku. Effort, on the other hand, decides how much time and work Claude puts into your task. Tariq, who works on the Claude code team at Anthropic, explains it with a simple comparison. If someone gives you 12 hours for a task, you probably assume they want you to just do it and try really hard by yourself. But if they give you just 1 hour for that same task, you probably assume they want a decent first version and expect to iterate from there. And effort works in a very similar way. So, at higher effort, Claude reads more, tests more, and double-checks its work more before it comes back to you. But as Lydia Holly from the Claude code team says, it means more than just thinking time. It also means it makes more decisions on its own, while at a lower effort, it would rather come back and ask you a question.
1:42 And that means effort is also a dial for how involved you want to be. On low, Claude keeps coming back to you, so you stay in the loop. And on high, Claude goes off and decides things for you, which is great when you're not around, but it also means Claude makes more choices for you, which is not always ideal. So, although many people assume the higher the better, this is really not the case and why it's so good to understand what effort level to actually use for which type of use case, which is exactly what I'll cover in this video.
2:08 Now, I've also built a mod that automatically analyzes your prompts and notifies you whenever you're likely using the wrong effort level for a task, which you can download and use for free by going to the first link in the description. You can just throw in the zip file into the Claude code tab and tell it to install this mod for you. And now, every time you throw in a prompt and it detects you're likely using the wrong effort level for it, it'll give you a pop-up with a suggestion on the new effort level, which you can then just click to apply, or you can even just decide to let this apply by itself. Now, I've been using this and it's been really useful. so, again, you can download it for free in the first link in the description. Now, there's one more interesting thing Anthropic found that's really important, which is that a bad prompt can't be fixed with a higher effort level.
2:50 Describing your task well is always important as the foundation. Starting with level one, low. Now, Anthropic's own rule for using low is "For when I want quick responses that are in the loop." Now, most people never use low, but low is actually a very useful effort level. So, you want to use low really in three situations. First, when you plan to iterate on something together with Claude, because you can react to a first version far faster and steer it in the right direction far faster. I, for example, now use low when prepping YouTube videos, because I really want to be in the loop here and steer it in the right direction. And it actually really speeds up the workflow and many people still think low is just dumb mode, but it's really not the case. And second, of course, you want to use low when a task is just simple. In Anthropic's own test, for example, he says that the same simple redesign took 1 minute on low and 28 minutes on max with a very similar outcome. And a third use case is when you're polishing something a higher effort level already made. So, for example, you do the hard first version on high and then switch to low for the small tweaks. So, let's say you create a report or a dashboard on high, and once you get a good initial version, you make the tweaks to the layout, the copies, and structure with low. And with the newest models, switching effort in the middle of a chat no longer makes Claude reread your entire conversation at full price, so it doesn't cost you anything extra. It will actually save you costs because you're not using high for small tweaks. So, some common tasks you probably want to use low for are brainstorming use cases, for example, something like give me 15 subject line options for this newsletter because you want options fast and you'll steer it anyway. also, quick rewrites or iterations, for example, make this client email friendlier and half as long because it's simple and clear. Then, sorting and tagging use cases, for example, tag these support tickets by issue type because in this case Claude is following a clear rules and volume matters much more than depth. then also, short summaries. Again, summarizing is a simple task for AI. And lastly, rough first drafts, for example, sketch an outline for next week's newsletter because again you will change it anyway. Now, one thing to keep in mind, low effort also means Claude does exactly what you ask and nothing more. So, if a task has several parts, Anthropic recommends giving Claude a clear or simple checklist. Now, level two is medium. Now, Tharic says medium is for most of his regular work, and Anthropic's official guide goes even further. For most tasks, you should use the model's default, and on Opus 5.5, that default is medium. So, you want to use medium really for most use cases when you can clearly define what one looks like and you'll review the result at the end once. Now, in practice, you probably want to use medium for the following tasks. Your everyday writing tasks, for example, write a follow-up email to Oscar from these call notes and our pricing. It's just a normal deliverable, so the default is more than enough. Also, repurposing use cases, for example, turn this YouTube video into a LinkedIn post because again the context is already in the source, so Claude doesn't really need to dig a lot deeper. Medium also works really well for anything around your meetings, for example, summarize this call with the decisions and the next steps. And the same goes for your standard reporting, like a weekly update or a client report that follows the same format every time.
5:56 Because Claude already knows what good looks like. And lastly, quick research, like getting up to speed on a client before a call. In short, medium should really be your default for most everyday tasks. And if a medium output isn't good, Lydia's advice is to not turn up the effort right away. Your first instinct should actually be to look at the context or prompt you gave it and only move up when Claude actually skipped a step or didn't check its work.
6:19 Now, before continuing, very quickly, if you enjoy this type of content, you'll really enjoy my AI Accelerator, where you'll get access to my 30-day Claude course that takes you from beginner to pro in less than 1 hour a day. We also have a full course on growing an AI business based on everything me and my team have learned over the last 3 years. We have unlimited one-on-one live AI consulting with AI experts and a big library of skills, templates, and resources across all business departments. So, if you want to dive deeper into AI, definitely check it out in the second link in the description.
6:47 Now, level three is high, and Tariq's rule is a high is for work where verification is important or there are edge cases. So, you basically want to use high when a mistake would actually cost you, like money, trust, or a decision made on wrong numbers. Second, you can also use it when a task has lots of edge cases, like checking the terms in a contract or the figures in a financial report. for example, in one of Anthropic's own tests, Claude had to analyze a data set, and on low, it just tried one approach and got it wrong every single time. And on high, it tried two approaches and noticed that the answers didn't match and figured out why, and then got it right four out of five times. And thirdly, you can also use high at the end to check and test your work. That's actually how Tariq works himself. He builds the first version on low, go goes back and forth on low, and only switches to high at the end to verify everything. So, some common tasks you might want to consider using high for, firstly, for checking numbers across different sources. For example, check this month's revenue report against Stripe and flag anything that doesn't match. Because in this case, one wrong number is easy to miss, so you want Claude to actually check every figure comprehensively. Now, secondly, for open-ended analysis, where it's really not clear yet what Claude should look into. For example, why did our churn go up last quarter?
8:03 because in this case, there's no set path, so you want Claude to actually explore a few angles to figure it out. Thirdly, you can use it for reviewing for risk. For example, review this contract and flag anything unusual in the payment terms. Again, contracts tend to have a lot of edge cases, and a mistake here can really cost you. you can also use it for fact-checking or for going through long documents like a 50-page report. Because in both these use cases, it's easy to miss something, and you want Claude to be thorough. Now, another good use case is the one from Tariq, right, which is verifying and testing outputs generated at lower efforts. For example, once your monthly report that you build on low looks right, you switch to high, and you say, "Check every number in this report against the source data and fix anything that's off." Because that way, you only pay for high on the final check. And then lastly, a good use case in my experience for high is building or fixing your skills. Because a skill runs the same way every time, so any mistake in it repeats every time, too. So, that's why you want Claude to test it properly before you rely on it. So, in short, you don't just want to use high because a task is long. Again, in Entropic's results, straightforward rule following work like operational tasks gain the least for more effort. So, it's much more about how much judgement and checking it needs. And then we have level four, extra high, and max. Now, I'm taking these two together because they're for the same kind of work.
9:23 And Tariq says he uses max basically only if I want zero input, meaning no input from you at all. Now, using these effort levels can mean Claude takes hours on a single task. So, you want to use these levels really only when all of the three following points are true. First, you just want to walk away for a couple of hours and not be there to answer any questions. Second, the job is very long and hard. And third, the work is critical, so the quality matters more than the time it spends and your usage limit. If even one of those isn't true, you probably want to use high instead because with these effort levels, Claude assumes it has to figure out everything on its own. So, it will make lots of decisions for you. It will double-check a lot and prompts become even more important at these higher effort levels because if you haven't described a task well, it can go on for hours and not even give you a good output. Now, some tasks you could consider extra high and max for are really long research tasks you want to run unattended. For example, go through every support call and ticket from the last year and give me a full report with the top reasons people cancel, a breakdown of our ideal customer profile, and a product development roadmap. I actually used this one my- myself. It's a really long job, but extremely valuable for decision-making. Now, a second use case can be a correctness-critical work. For example, review this 50-page agreement before we sign it and flag everything that could hurt us. Now, you're preventing a real risk here, so the extra cost might be worth it. And lastly, when you want what Atharic calls Claude's best one-shot. For example, create our investor pitch deck from our metrics, a strategy doc, and past decks.
10:56 Now, this case it's high stakes, so you want the most complete version on the first try, and then you can always iterate at a lower effort. But again, just keep in mind that Anthropic's own docs say that on most workloads, max specifically adds significant cost for relatively small quality gains. And on some tasks, it can even lead to overthinking where Claude keeps going and the results actually get worse. Now, extra high is basically a cheaper version of max, so always start with extra high and only go to max when it's really critical. So, for most people watching, max should be very rare. So, in short, whenever you're not sure which level to use, you can just ask yourself the following four questions in this order and stop at the first yet. Will you go back and forth with Claude on this or is it a simple task? Then you want to use low. Would a hidden mistake cost you like money, a client, or a wrong decision? Then use high. Are you handing off a long, hard, and critical job and walking away? Then use extra high or max. And for everything else, then just use medium, which is the default on Opus 5.5. And if you're unsure, again, just use medium. So, the main takeaway really is effort isn't a quality dial. It's a dial for time and how involved you want to be. So, you only pay for more effort when a task actually requires it. Again, if you want to download the free mod that automatically applies the effort levels to your prompts, you can download it in the first link in the description. And if you want to learn more about my AI accelerator and my full Claude courses, you can check out the second link in the description below. Thank you so much for watching. If you got any value out of it, I highly appreciate a like and a subscribe. It really does help me. And if you want to learn more about working with Claude, you can check out the video here above.