Big Idea

The argument in one line.

Every skill, connector and memory file you add loads into every Claude chat, so a setup you never prune quietly gets slower and costlier, and a weekly automated audit fixes it.

Who This Is For

Read if. Skip if.

READ IF YOU ARE…

You use Claude Code or Cowork daily and have accumulated dozens of skills, connectors and chats you never clean up.
You keep a second brain or CLAUDE.md files and suspect some of the rules in them are outdated or contradict each other.
You are hitting Claude usage limits and want to find out how much baseline context your setup burns before you type anything.
You run scheduled tasks or routines and have no idea which ones are failing or still worth running.

SKIP IF…

You use Claude occasionally with a handful of chats and no custom skills, connectors or scheduled tasks.
You want a deep technical walkthrough of how the skill is built, since this video only shows what it does and the results.

TL;DR

The full version, fast.

As you add skills, connectors, memory notes, context files and scheduled tasks to Claude, a large part of that setup loads into every chat, making Claude slower, more confused and more expensive. Ben AI built a free Detox skill that sends five sub-agents to audit your interface, skills, memory and context files, connectors and scheduled tasks, then lays every proposed fix on one reviewable page. You approve, and it archives, deletes, regroups or rewrites automatically. On his own setup it found 239 items and saved about 15,000 tokens per chat. He recommends running it weekly or biweekly as a scheduled task.

Free for members

Chat with this breakdown — free.

Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.

Create a free account →

Chapters

Where the time goes.

00:00 – 01:08

01 · The problem and the promise

Setups bloat with skills, connectors, tasks and chats, and the Detox skill is introduced.

01:08 – 02:53

02 · Why keep it clean

Four reasons: your own productivity, better outputs, context and usage limits, and outdated rules for older models.

02:53 – 04:24

03 · How the Detox skill works

Five sub-agents audit five areas, produce one approval page, and apply fixes on Begin optimizing.

04:24 – 06:00

04 · Interface

Chats, pinned chats, artifacts and mods get archive, group or switch-off suggestions.

06:00 – 08:28

05 · Skills

Finds unused, duplicate and overlapping skills and can turn off auto-triggering.

08:28 – 11:21

06 · Memory and context

Audits memory and global instructions, project files and second brains, and CLAUDE.md files.

11:21 – 13:37

07 · Connectors

Lists connector cost and last use, flags expired logins and risky write access, and suggests Composio or CLIs.

13:37 – 15:36

08 · Scheduled tasks

Shows failing, unused and over-modeled tasks and recommends pausing or moving them to Sonnet.

Atomic Insights

Lines worth screenshotting.

Skill descriptions, connectors, memory and CLAUDE.md files load into the context window of every chat, even when you never use them.
When a better model ships, the first thing to try is removing instructions, because rules written for older models can hold new ones back.
Ben had 156 skills loaded and used only 8 in 30 days, and the unused list alone cost about 10,000 tokens in every chat.
Adding disable-model-invocation to a skill makes it fire only on its slash command and stops its description loading into every chat.
25 connected connectors cost about 16,000 tokens per chat, and switching off the 19 unused ones saved 11,000.
CLIs are more context efficient than MCP connectors because they do not add a list of tools to every chat.
A single strategy file had six stacked copies of itself with contradicting pricing information, which Claude would read as truth.
Of 23 scheduled tasks and routines, only 6 were used consistently, 1 was failing and 18 still ran on older models.
Simple scheduled tasks like summaries usually run fine on Sonnet, so keeping them on Opus is wasted usage.
Context files and second brains rot over time: duplicate notes, contradictions and notes nothing links to pile up with no way to track them by hand.
A scheduled audit that produces a checklist is easier to sustain than a cleanup habit that depends on remembering.

Takeaway

A lean setup beats a bigger setup

WHAT TO LEARN

Everything you add to Claude loads into every chat, so regular pruning of skills, connectors, memory and tasks saves tokens and improves output quality.

02Why keep it clean

Skill descriptions, connectors, memory and CLAUDE.md files load before you type, so unused ones quietly burn usage in every single chat.
When a stronger model ships, try removing instructions first, because rules written for older models can hold the new one back.

04Interface

Archive instead of delete when unsure with chats, since archived chats can be brought back and the cleanup stays low risk.

05Skills

Audit skills for overlap, duplicates and accidental firing, and use disable-model-invocation so rarely used skills run only by slash command.

06Memory and context

Contradictory or duplicated rules in memory and CLAUDE.md files make Claude follow stale instructions, so review them against what you actually asked for recently.
Second brains need indexes, working links and no duplicate or orphaned notes, or the extra context starts hurting output quality.

07Connectors

Score each connector by last use and context cost, switch off unused ones, and prefer one login layer like Composio or a CLI over many separate connectors.
Check connectors for risky write access, such as Gmail being able to send email when Claude only needs to read.

08Scheduled tasks

Review scheduled tasks for failures, unread outputs and model choice, and move simple recurring jobs from Opus to Sonnet.
Put the audit itself on a weekly or biweekly scheduled task so cleanup becomes a checklist you approve rather than a chore you remember.

Glossary

Terms worth knowing.

Harness: Everything around the model itself: skills, connectors, memory, instructions and tools. It goes out of date quickly as new models ship.
Baseline context: Everything that loads into a chat before you send a message, such as skill descriptions, connector tool lists, memory and CLAUDE.md files.
Skill: A saved how-to file that teaches Claude to do a task your way.
Connector: An integration that gives Claude access to an outside app, adding its tool names and instructions to every chat.
Composio: A meta connector that gives Claude access to many apps through a single login and handles reconnections automatically.
CLI: A small command line tool Claude can run instead of a connector, without adding a tool list to every chat.
Routine: A scheduled Claude Code task that runs in the cloud, so it works without your laptop being open.

Resources

Things they pointed at.

02:40productClaude Detox Skill (free) ↗

08:16productAI Accelerator ↗

12:35toolComposio

13:10toolPrinting Press (CLI builder) ↗

Quotables

Lines you could clip.

02:28

“The best cloud users also make sure their harness, which is essentially everything around the model, stays clean.”

Short, reframes maintenance as a skill→ newsletter pull-quote↗ Tweet quote

03:46

“It found 239 things to clean up and saved around 15,000 wasted tokens for every single new chat I start.”

Concrete before and after numbers→ TikTok hook↗ Tweet quote

07:35

“I had 156 skills loaded and I'd only used eight of them in the last 30 days.”

Relatable, specific, no setup needed→ IG reel cold open↗ Tweet quote

The Script

Word for word.

Read-along

Don't just watch it. Burn it in.

See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.

The more people use Clod, the more they add to it. More skills, connectors, scheduled tasks, context files, artifacts, and hundreds of chats. That's a good thing because that's how Clod gets more useful.

But barely anyone actually maintains, cleans, or optimizes this entire setup. So it becomes impossible to find old chats. We have dozens of skills we never use.

Clod still uses outdated roles and context that changed months ago. And scheduled tasks quietly fail. And besides this becoming an overwhelming experience every time you open up Clod, a big part...

this bloated setup also loads into every single chat. So instead of Cloth getting better, it gets slower, messier, and more expensive. So in this video, I'll show you how to keep your entire Cloth setup clean with one skill that I built that you can download for free in the first link in the description.

I'll show you how this detox skill cleans up and optimizes everything from your Cloth interface, chats, projects, artifacts, skills, scheduled tasks, and even your context or your second brain if you have one, and show you how you can automatically run it every week to get the And if you're new here, I'm Ben. I'm a three -time founder and now run an AI agency and one of the largest AI communities where I help professionals and business owners get real results with AI.

Now, before I show you the skill, why does maintaining your cloud setup actually matter so much? Even Boris Cherny, who created Cloud Code, said himself, my setup might be surprisingly vanilla. And it's because a clean setup is important because of four big reasons.

First, it's for you, of course. The fewer chats, skills, and files you have to dig through, the more productive you'll get. It's basically like working at a clean desk instead of a messy one.

Second, it will lead to better outputs. Every rule, memory file, skill, and connector is something Clot takes into account. So when they're outdated, they contradict each other or are irrelevant, Clot gets confused, follows an old rule, or uses a skill nobody asked for.

Thirdly, it will impact your context and usage limit. Now, many people don't know this, but a big part of this setup loads into every chat before you even type a message. your skill descriptions, your connectors, your memory and Cloud .MD files.

It all loads into the context window, even when you don't use it, which burns through your usage far faster and affects Cloud's performance. And lastly, another big reason for this is that every time a new, better model comes out, a lot of the rules people wrote for the older ones aren't needed anymore. And some even hold these new models back.

And that's why Tariq from the Cloud Code team also says that when a better model comes out, the most important first thing you should try is to remove instructions. And with new models coming out every few months, Tariq himself says that, harnesses go out of date very quickly.

So while building and adding is the exciting part, of course, the best cloud users also make sure their harness, which is essentially everything around the model, stays clean. optimized and up to date but because of course that's boring to do i've made it really easy to do with our detox scale which you can just download by going to the first link in the description uploading it to your cloud account and running it with slash detox now you can run this scale either in the cloud code tab or in the co -work tab in the co -work tab it will have some limitations compared to how it works in the cloud code tab but it will still be extremely valuable so what does it do it essentially audits cleans and fixes five parts of your cloud setup your interface, your skills, your memory and context files, your connectors, and your scheduled tasks.

So when you run it, it sends out five sub -agents, one for each part of your setup. So every part gets properly audited at the same time without filling up your context window. It then puts everything it found into one big page.

On that page, it will lay out everything it found across each part of your setup. For every single item inside of it, Cloud will suggest a fix like archive or delete or a change. then still change what happens to each item or leave a comment on it and if you're happy with the suggestions you can just approve them all by just clicking begin optimizing and Cloud will organize and clean up everything for you automatically.

For example, when I ran it on my own setup, it found 239 things to clean up. It deleted some old sessions, grouped sessions, spotted failed connectors, fixed my context files in Cloud .MD, and much, much more. And after it was finished, it also saved around 15 ,000 wasted tokens for every single new chat I start.

I now run this detox skill automatically on a scheduled task every two weeks. to make sure my cloud setup is cleaned up and optimized consistently, which you can just do by setting up a scheduled task, say run the detox scale and setting the schedule for every week or every two weeks. Now the scale will do all of this for you, but it really helps to understand what it actually audits and fixes.

So let me go through it, starting with the interface. Now by interface, I mean everything you see and click in cloud, your chats, pin chats, projects, artifacts. like your dashboards and slides, and your mods, which are basically small add -ons that change how Cloud Code looks and behaves.

So these are the first things the detox skill looks at because cleaning this up already takes away a lot of that overwhelm. So firstly, it'll go through your chats to find all the ones you don't need or use anymore, like test chats, one -off questions, and personal chats. It already suggests what to do with each one.

You can still decide if you want to archive or delete them, and an archive chat you can always bring back. Then it also suggests groups. Groupings for all the chats that are left.

For example, in my case, it grouped chats by YouTube videos. And once I approve, it moves them into these groups automatically. And this is honestly one of my favorite things this skill does because I also always had a problem with finding my chats.

But keep in mind this grouping feature only exists in the Cloud Code tab. You'll also look at your pinned chats. If you're actually still using them, you'll suggest unpinning old ones and pinning chats you come back to frequently.

It also looks at all your artifacts like your dashboards, designs, docs, slides, and motion graphics. It finds the ones you haven't used in a while with a link to each so you can check before changing or deleting anything. And lastly, also check your mods if you've already been using them.

First, it checks if you have any. And if you do, it lists each one, whether it's switched on or off, and if you've actually been using it, and then suggests the ones to switch off. Now, honestly, cleaning up your interface like this every week or two has been a real game changer for my productivity.

So this alone is already extremely valuable, but it also solves another big problem I've been facing, which is ending up with dozens or even hundreds of skills. that I haven't used in months.

Now, if you're still unfamiliar, a skill is basically just a saved how -to file that teaches Glot how to do a task your way. And once people start building them, they often end up downloading and building lots of them, which creates a couple of problems. Firstly, some skills will overlap or do similar things and you don't know which ones to use anymore.

Second, you often end up with duplicate ones and different versions. Third, some skills might fire by accident in tasks where you don't want them to be fired. and lastly every skill costs context and tokens because the descriptions of every skill loads into every chat even if you don't use these skills so the detox skill basically first finds every skill you have where it lives so you can find duplicate ones then it checks which skills you haven't used in a while which ones overlap and which ones fire when you didn't ask for them you don't get one clean table sorted by category with when you last used each skill how much context they cost and whether clods suggest keeping or removing it.

You can still check. And once you approve, it'll clean your skill set up instantly. And for skills that keep firing by accident, it can even add one line to the top of the skill file called disable model invocation, which means we can essentially turn the auto trigger off for the skill.

So it can only fire when you use the slash command. This is honestly my preferred way of using skills. And it's also good because it stops the description from loading into every chat, which saves you tokens and context too.

Now, in my case, for example, I had 156. scales loaded and I'd only used eight of them in the last 30 days. It also found 10 scales sitting in more than one place.

and 15 that basically do the same job, like four old LinkedIn writer skills next to the ones I actually use. In Cloud Code, this unused skill list alone was costing me about 10 ,000 tokens in every chat. Now, the next and probably most important part of fixes is your memory and your context.

But before showing you what it does there, 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 Cloud course that takes you from beginner to pro in less than one hour a day. You also get access to a full course on... growing an AI business based on everything me and my team have learned over the last three years and access to unlimited one -on -one live AI consulting.

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 below. Now, memory and context are, of course, vital to get the most out of Cloud, but this is also the part that can get messy and outdated really fast, which can significantly affect your output quality.

So the Detox Scale goes through three things here. First, your built -in memory and global instructions. So Cloud basically saves memory nodes on its own based on your feedback and preferences.

which you can find in your settings under the memory tab where you can also find your global instructions which is basically a file with instructions cloud reads at the start of every single chat and over time these go out of date right people change how they want something done but an older note still says the opposite or two notes say something completely different so the detox scale reads them plus your chats from the last 30 days to see what you've actually asked for lately and flags rules that are outdated, duplicated, or contradict each other.

Then again, it suggests what to fix. And once you approve, it can automatically update your rules and memory files. Second, you also do a full audit and optimization of the files inside of your projects and folders.

Now, if you've worked with folder access, cloud projects, or even might have a second brain where you save your context and memory notes, you can easily end up with dozens, hundreds, or even thousands of files and notes, which is, of course, extremely valuable. This cloud can give you far better outputs with this context.

But as this grows, it will start to bloat. You'll have duplicate nodes, nodes that contradict each other, and nodes nothing links to anymore. And there's really just no way to keep track of all of that yourself.

So the detox scale goes through all of those files in your folders and your projects and audits everything based on best practices from Anthropic and Andrej Karpathy. It fixes broken links, cleans up duplicates and outdated nodes, flags nodes nothing points to, and makes sure every bigger folder has an index file. So it uses around...

16 frameworks and best practices for optimizing context or knowledge basis. to make sure all of this stays relevant and working well. And thirdly, maybe most importantly, it audits and optimizes your cloud .md or agents .md for each project or folder.

Now the cloud .md or agents .md, if you use multiple AI for providers, essentially tells cloud where everything lives inside of these folders. Now the cloud .md or agents .md is of course one of the most important files for these AI agents, but we do have to optimize them regularly according to what lives in our folders, our roles, and our preferences.

So again, the details TalkSkill checks that your Cloud .MD still points to the right places, flags important folders it doesn't mention, offers a cleaner, shorter rewrite when it has grown too long. and suggest creating one where it's missing.

It does this for every project you're actively working on over the last 30 days. In my own setup, for example, it found seven rules that contradicted each other. It found that my main cloud .md still pointed to a tasks folder that never existed.

And my strategy file, the one cloud uses for our pricing, had six copies of itself stacked on top of each other with contradicting information. Now, the next thing it fixes are your connectors. Now, if you've been using them, you probably have noticed that Logins often expire or you have to reconnect them.

Cloud can also have full write access to tools where it only needs to read. the other way around also many people keep connectors they don't actually use anymore some might even overlap like a custom mcp and an official connector that later came out connectors are also one of the biggest parts of your baseline context which basically means everything that loads into a chat before you even send a message because every connector adds its tool names and instructions to every chat even when you don't use it so the detox scale lists every connector with when you last used it and how much context it costs you it then suggests switching off the unused ones and keeping just one of each duplicate.

It flags expired logins and risky write access. Now, Cloud can't actually reconnect the connector for you, so you can do that yourself based on Cloud's suggestions. Then it even goes one step further.

For every connector, it will also check if the tool is also on Composio, which is basically a meta connector that gives Cloud access to a lot of apps at once with just that Composio connector. If it is, it will suggest switching to Composio because with Composio, you can essentially do one login on all your softwares and give Cloud access to...

all of them with just the Composio connector. Composio even manages the reconnections automatically for you. Now, I'm not being paid to say this.

I just really like the tool Composio. So I highly recommend checking it out. And if a connector isn't on Composio, it might also suggest building a CLI.

A CLI is basically a small command line tool Cloud can run instead of a connector. And Entropic's own docs say CLIs are more context efficient than MCPs, which are connectors, because they don't add a list of tools to every chat. Now, you can really easily build your own CLIs with something called the Printing Press, an open source tool that builds a CLI from an app's API docs.

I'll leave the link to the Printing Press in the description below too. But in my own setup, I had 25 connectors connected, costing me about 16 ,000 tokens in every chat, 19 of them I hadn't used in the last 30 days, and just switching off the unused ones saved me 11 ,000 tokens per chat. It also found five services I had connected twice, three I could move to Composio, and seven connectors with risky write access.

like Gmail being able to send emails. And then the last part is the easiest one to forget because it all runs in the background, your scheduled tasks and routines. Now, one big problem I have faced with these scheduled tasks or cloud code routines, which is essentially the same as a scheduled task, but runs in the cloud.

So it can run without your laptop being open is that although they're very valuable, of course, many of them quietly fail because a connector fails, for example, or a prompt is outdated. And many that I've set up in the past, I don't actually use anymore. or even check the outputs of anymore.

And because they run in the background, it's easy to forget about them. But of course, they do eat away at your plan's usage every single time they run. So the Detox Scale lists every scheduled task, both in co -work and in a co -tab, with when it last ran, whether it failed, and which model it runs on.

Now, it will show failing tasks first with the error, like a login that needs reconnecting. For tasks whose output nobody actually reads anymore, it suggests pausing these scheduled tasks. And it also checks which ones are simple enough to run on Sonnet instead of Opus, for example.

Entropic's own advice is to keep Opus for complex multi -step work and for simple tasks like summaries, Sonnet is usually more than enough. So it can suggest switching some of these scheduled tasks to cheaper models. In my setup, for example, I found 23 scheduled tasks and routines.

Only six of them I was actually using consistently. One was failing and 18 were still running on older models. And four of them were simple enough to run on Sonnet.

And once you approve, it can pause or switch them off automatically for you. Again, you can download the Detox Scale for free in the first link in the description. I highly recommend putting it on a weekly scheduled task to audit the setup for you.

And then all you have to do is go through the checklist whenever it's ready. And again, if you want to get access to unlimited one -on -one AI consulting, my full 30 -day beginner to advanced cloud... course, hundreds of skills and resources like these and Q &As with me and my team.

Definitely check out my AI Accelerator in the second link in the description. As always, thank you so much for watching. If you enjoyed this video, I highly appreciate a like and a subscribe.

It really does help me. And if you want to learn more about cloud, you can check out the video here above.

The Hook

The bait, then the rug-pull.

Claude gets more useful the more you add to it, but almost nobody prunes what they added. Ben shows how a bloated setup makes every chat slower and pricier, and demos a free skill that cleans it all up.

Frameworks