Source: raw/This_Skill_Instantly_10x_es_Every_Claude_Output.md

Creator: Ben AI · URL: https://www.youtube.com/watch?v=d2XiTDzoizc · Duration: 14:19 · Platform: YouTube · Uploaded: 2026-07-02

Many people keep returning to a few huge chats (a copy-review chat, a per-video YouTube chat) because starting fresh means re-explaining all the rules and context. Ben’s answer is a habit plus a skill: check /context to see how full the chat is, and when it is too full, run his free refresh skill. The skill writes the chat’s state, rules and context-file list to disk and produces a starter prompt for a new chat. He built it for Cowork knowledge work, crediting Matt Pocock’s engineering-focused handoff skill as the inspiration (Matt Pocock’s skills).

Key Takeaways

  • Ben’s context zones (his rule of thumb on a 1M-token Opus window, not an Anthropic figure):
    • 0 to 200K tokens: best output.
    • 200K to 350K: still useful, starting to drift.
    • 350K to 400K: noticeably worse.
    • 400K+: the “you’re right to push back” zone. Start a new chat.
  • Use /context to see where you are. Claude Code in the terminal can show token use; Ben says the Claude desktop app (Cowork and the Code tab) has no on-screen counter, but /context works there and breaks down what used the tokens.
  • When to check: long sessions; chats you keep coming back to; sessions with heavy connector, MCP or research use (Ben finds these the biggest token users); and when outputs get worse, you start repeating yourself, or you feel frustrated.
  • Why not /compact: Ben names four problems for his use. It only works in Claude Code, not Cowork. Its summary is too short (he says about 50 to 60K tokens), so you re-explain things. The small rules you added during the chat get lost. And each context file you fed in shrinks to a one-line summary, so you have to feed them in again.
  • What refresh does differently. It asks the goal of the next session first. It writes several files into the working folder (confirmed outline, intro, research findings, decisions and rules). It then writes a starter prompt that states the goal, tells the new chat to read those files plus the original context docs (strategy, ICP, brand, transcripts), and lists where things stand, what worked, what didn’t, and corrections made.
  • It costs more tokens than /compact: Ben puts the fresh session at around 100K tokens after the refresh. He says that is worth it to avoid re-prompting.
  • Second use: switching tasks. If a YouTube-ideation chat turns into writing a newsletter, refresh into a new chat with the goal “write a newsletter from this context” rather than polluting the original chat. In his demo, the new chat read the files and then invoked his newsletter skill.
  • Always select a working folder in Cowork or Claude Code. The refresh files and your context docs need a shared folder the next chat can read.

Try It

  1. In your longest-running chat, type /context and note the total and what used most of it.
  2. If it is past roughly 200K and output quality has slipped, run a handoff. Without Ben’s skill, ask: “Before I start a new chat: ask me the goal of the next session, then write files to this folder capturing current status, decisions, every rule or correction I gave in this chat, what worked and what didn’t, and the full list of context files you read. Then give me a starter prompt that tells the next chat to read all of those first.”
  3. Open a new chat with the same folder selected, paste the starter prompt, and confirm it reads the files before it does anything else.
  4. Compare the first output in the new chat against the last output in the old one.
  5. Once it works, save the instruction as a skill so it runs the same way each time.

Implementation

Tool/Service: Ben’s refresh skill (free download from his video description); /context and /compact are built in.

Setup: Install the skill in Cowork or Claude Code; work inside a selected folder so the handoff files persist.

Cost: A larger opening context in the new chat (Ben’s estimate is about 100K tokens) in exchange for fewer re-explanations.

Integration notes: For persistent rules that should outlive any one chat, a skill or CLAUDE.md is the better home than a handoff file.^[inferred] Compare Anthropic’s own advice in Context Management in Claude Code (/compact with focus instructions, /clear to start fresh) and Claude Code Best Practices (“The context window is the most important resource to manage”).

Open Questions

  • The zone thresholds are Ben’s experience on Opus 4.8 at 1M context. They are not measurements, and newer models may hold up longer.
  • “/compact only works in Claude Code”: not checked against the current desktop app, which merged chat and Cowork in September 2026 (Claude Cowork).
  • The 50 to 60K /compact summary size is Ben’s figure; Anthropic does not state one in the wiki’s sources.