Source: raw/This_ONE_Skill_Will_10x_Your_Claude_Setup.md
Creator: Ben AI · URL: https://www.youtube.com/watch?v=k6JWDYvHr3U · Duration: 15:36 · Platform: YouTube · Uploaded: 2026-11-06
Ben AI built a free skill, /detox, that audits five parts of a Claude setup: the interface, skills, memory and context files, connectors, and scheduled tasks. It sends one subagent per part, collects every finding on one review page with a suggested action, and applies the actions you approve. His case is that much of a setup loads into every chat before the first message, so unused skills and connectors cost tokens and stale rules confuse the model. The skill is a download linked from his video description and was not reviewed here; this article records what he says it does and the numbers from his own setup. The auto-captions render Claude as “Clutter” and “cloud”; corrected here.
Key Takeaways
- Four reasons he gives for cleaning up:
- Fewer chats, skills and files to dig through.
- Better output: outdated or contradictory rules, memory files, skills and connectors make Claude “follow an old rule, or use a skill nobody asked for.”
- Usage: skill descriptions, connectors, memory and CLAUDE.md files load into every chat even when unused.
- New models need fewer rules. He relays Thariq of the Claude Code team saying the first thing to try with a better model is to remove instructions, and that harnesses go out of date quickly. The wiki’s first-party record of the same idea is Anthropic cutting over 80% of Claude Code’s system prompt for Opus 5 (New Rules of Context Engineering for Claude 5).
- He also quotes Boris Cherny: “my setup might be surprisingly vanilla.” The quote is not checked here.
- How it runs. Upload the skill and type
/detox, in the Code tab or the Cowork tab. In Cowork it has “some limitations” he does not list. Five subagents audit the five areas at once “without filling up your context window.” - One review page. Every item gets a suggested action (archive, delete or change). You can change any action or comment on an item, then click “begin optimizing” to apply all of them.
- His first run found 239 items and, he says, saved about 15,000 tokens on every new chat.
- He reruns it on a schedule. A scheduled task with the prompt “run the detox skill”, every two weeks for him. Every one or two weeks is his advice.
The five audits
1. Interface
- Chats: finds test chats, one-off questions and personal chats, and suggests archive or delete. An archived chat can be brought back.
- Grouping: suggests groups for the remaining chats (his were grouped by YouTube video) and moves them once approved. He notes grouping exists only in the Code tab.
- Pinned chats: suggests unpinning old ones and pinning chats you return to often.
- Artifacts: lists dashboards, designs, docs, slides and motion graphics you have not opened in a while, each with a link so you can check before deleting.
- Mods: lists each mod, whether it is on and whether you use it, and suggests which to switch off. See Claude Mods.
2. Skills
- Finds every skill and where it lives, so duplicates show up, then checks which are unused, which overlap and which fire when you did not ask.
- Output is one table by category with the last-used date, context cost, and a keep or remove suggestion.
- For skills that fire by accident, it adds
disable-model-invocationto the top of the skill file, so the skill runs only from its slash command. Ben says this also stops the description loading into every chat. ^[ambiguous] The wiki records the first part as documented and the second as unconfirmed; see Skill Authoring Best Practices. - His numbers: 156 skills loaded, 8 used in the last 30 days; 10 sitting in more than one place; 15 doing the same job, such as four old LinkedIn writer skills. In Claude Code the unused skills cost him about 10,000 tokens per chat.
3. Memory and context
- Built-in memory and global instructions (Settings, Memory tab). It reads them with your last 30 days of chats and flags rules that are outdated, duplicated or contradict each other, then updates them once approved.
- Files in projects, folders or a second brain. It applies what he calls “around 16 frameworks and best practices” from Anthropic and Andrej Karpathy. It fixes broken links, removes duplicate and outdated notes, flags notes nothing links to, and makes sure every bigger folder has an index file. He does not list the frameworks.
- CLAUDE.md or AGENTS.md for each project active in the last 30 days. It checks the file still points to the right places, flags important folders it does not mention, offers a shorter rewrite when it has grown too long, and suggests one where it is missing.
- His findings: 7 rules that contradicted each other; a main CLAUDE.md pointing to a tasks folder that never existed; and a pricing strategy file with “six copies of itself stacked on top of each other with contradicting information.”
4. Connectors
- Lists every connector with when you last used it and its context cost. Each connector adds its tool names and instructions to every chat, so connectors are “one of the biggest parts of your baseline context.”
- Suggests switching off unused connectors and keeping one of each duplicate, such as a custom MCP next to a later official connector.
- Flags expired logins and risky write access. Claude cannot reconnect a connector for you.
- Composio check. If a tool is on Composio, it suggests moving it there: one login reaches many apps through a single connector, and Ben says Composio handles reconnections. He says he is not paid to recommend it.
- CLI suggestion. Otherwise it may suggest building a CLI, for example with Printing Press. Ben cites Anthropic’s docs as saying CLIs are more context-efficient than MCP connectors because they do not add a tool list to every chat; he does not say which page.
- His numbers: 25 connectors costing about 16,000 tokens per chat; 19 unused in 30 days; switching those off saved 11,000 tokens per chat; 5 services connected twice; 3 that could move to Composio; 7 with risky write access, such as Gmail being able to send email.
5. Scheduled tasks and routines
- Lists every scheduled task in the Cowork and Code tabs, and cloud routines, with the last run, whether it failed, and which model it uses. Failing tasks come first, with the error (for example a login that needs reconnecting).
- Suggests pausing tasks whose output nobody reads.
- Suggests moving simple tasks from Opus to Sonnet. He relays Anthropic’s advice to keep Opus for complex multi-step work, with Sonnet usually enough for summaries.
- His numbers: 23 scheduled tasks and routines; 6 used consistently; 1 failing; 18 on older models; 4 simple enough for Sonnet.
Try It
- Measure first. Open the context card in the session rail on a fresh chat and note how much is used before your first prompt (see The Revamped Claude Desktop App).
- Run the built-in checks before installing anything.
/skill-doctorreports unused skills and their cost,/doctorchecks settings and CLAUDE.md, and/checkupcleans unused skills and MCP servers after confirming (CLI Reference). These cover parts of audits 2 to 4. - Switch off connectors you have not used in 30 days, and keep one of any pair that reaches the same service.
- Add
disable-model-invocation: trueto skills you only ever call by name. - Review scheduled tasks: pause the ones whose output you no longer read, and move summaries and other simple jobs to Sonnet.
- If you install his skill, read its SKILL.md first. It archives and deletes, so check the review page before “begin optimizing”.
- Schedule the audit every one or two weeks.
Implementation
- Tool/Service: Ben AI’s free
/detoxskill (download link in the video description). - Setup: upload the skill to Claude and run
/detox; the Code tab gives the full audit, the Cowork tab a limited one. - Cost: the skill is free. The audit runs five subagents, so it uses plan usage; he gives no figure.
- Integration notes: pair it with a scheduled task for the recurring run. For CLAUDE.md alone, the doc-cleanup skill covers a similar check.
Open Questions
- The skill itself was not reviewed. Which “16 frameworks” it applies, and what it changes without asking, are not shown in the video.
- His token figures do not add up as stated. Switching off connectors saved 11,000 tokens and unused skills cost about 10,000, yet the total saving per chat is given as about 15,000. He may not have removed every unused skill; he does not say.
- Does
disable-model-invocationremove the description from context? Still unconfirmed; see Skill Authoring Best Practices. - Which Anthropic page says CLIs are more context-efficient than MCP? Not named. The Printing Press article records a separate creator claim that MCP used 35 times more tokens than a CLI on one task.
- Cowork-tab limitations are mentioned but not listed.
Related
- The Revamped Claude Desktop App — the session rail’s context card and the
/skill-doctorand/doctorresults other creators report. - Claude Code CLI Reference —
/checkup,/doctorand/skill-doctor. - New Rules of Context Engineering for Claude 5 — Anthropic’s own case for removing instructions as models improve.
- Skill Authoring Best Practices —
disable-model-invocationand skill descriptions. - CLAUDE.md Context Rot Cleanup — a narrower skill for stale CLAUDE.md content.
- Printing Press — building a CLI from an app’s API docs.
- Claude Connectors — Composio and other middleware connectors.
- Claude Code Scheduled Tasks — the mechanism for the recurring audit.