Claude Code Best Practices: 7 Levels to Double Your Results

By Beau Johnson··8 min read

Quick answer

To get more out of Claude Code, stop being the loop yourself. Give it memory with a short CLAUDE.md, eyes with the Chrome DevTools MCP, a check it can loop on with /goal, guardrails with hooks, and clean context. Once all that is in place, run parallel agents in worktrees.

Anthropic looked at about 400,000 Claude Code sessions and found that people who knew how to use it got about double the verified results of beginners. Same tool, same model. In this video I show how the best builders get there, with real demos on a real project: a rough AI-made landing page for a habit app called Pulse. It's for anyone who uses Claude Code and keeps getting stuck fixing its mistakes by hand.

What I cover in this video

  • 0:00: Same tool, double the results (the Anthropic study)
  • 0:45: Stop being the loop
  • 1:31: Level 1, giving Claude memory with CLAUDE.md
  • 3:01: Level 2, giving Claude eyes with the Chrome DevTools MCP
  • 3:47: Level 3, skills and the Impeccable design audit
  • 4:34: Level 4, /goal, /loop, and auto mode
  • 6:05: Level 5, hooks as guardrails (and how Claude got around mine)
  • 8:24: Levels 6 and 7, Ponytail, then parallel agents with worktrees

The core idea: stop being the loop

Most people use Claude Code like this: you ask, it writes code, you check it, you find the bug, you tell it, it fixes it, and you check again. That makes you the bottleneck. Boris Cherny, who leads Claude Code at Anthropic, says his job now is to write loops, not prompts. He's also said that giving Claude a way to verify its own work "will 2-3x the quality of the final result." Everything in this video is about getting you out of that loop safely. (In the Anthropic study, verified success was 15% for novices and 28–33% from intermediate up.)

Levels 1–3: memory, eyes, and skills

Give Claude memory with CLAUDE.md

Every session, Claude starts from zero. CLAUDE.md is a file it reads at the start of every session, and you don't have to write it yourself. Run:

/init

It took 24 seconds. Claude read the project, found the commands and the layout, and even caught a gotcha: the page needs a server to run because of how the scripts load. The habit that pays off over time: whenever Claude makes a mistake, tell it to update your CLAUDE.md so it doesn't make that mistake again. Keep the file short, though. Anthropic's docs say under 200 lines, because a bloated CLAUDE.md gets ignored. If a rule has to happen every single time, it belongs in a hook instead.

Give Claude eyes with a browser

Claude can write a page but can't see it, so it guesses. I added the Chrome DevTools MCP server, which takes one command, and then prompted:

Open the page in the browser, check the console, and fix anything that's broken.

It opened a real browser, took a screenshot, read the console, found a script error that fired on every page load, fixed it, and checked again, all in 48 seconds. Then, without me asking, it clicked the yearly pricing toggle and caught a price of $160 followed by about 14 zeros and a 3. The other MCP server the pros recommend is Context7, which feeds Claude the current docs for your libraries so it stops writing code for old versions.

Add skills like Impeccable

A skill is a folder of instructions Claude only loads when it needs it, so it costs almost nothing until it's used. Impeccable (about 85,000 GitHub stars) is a design skill that knows the tells of an AI-made website: the purple-to-blue gradient, the emoji in a rounded square, cards inside cards. I installed it as a plugin (two commands) and ran:

/impeccable audit

My page scored 5 out of 20, rated critical. The audit flagged 14 AI template patterns, text that fails accessibility contrast, and a page almost twice as wide as a phone screen. Then I ran /impeccable polish, and about 3 minutes later it had the same product and sections but looked like a real company built it.

Levels 4–5: loops, auto mode, and hooks

The pricing had three tests, and all three were failing. Instead of going back and forth, I typed:

/goal npm test passes with all three pricing tests green

/goal keeps Claude working turn after turn until a check confirms the condition is true. It stops when the tests pass, not when Claude thinks it's done. 41 seconds later, all three passed. There's also /loop, which reruns a prompt on a timer while your session is open, and /schedule, which runs a task in the cloud on a schedule even with your laptop closed.

Auto mode has been the default starting mode since v2.1.283 (September 25, 2026). A separate safety model approves the safe actions so you aren't clicking yes 100 times. More power only works if you set rules, though. I mention the story about big US banks finding AI agents inside customer accounts because people handed their logins to AI assistants.

Hooks are rules that run every time, no matter what the model decides. I had Claude add a hook blocking it from reading or editing my .env file. Then I tested it by asking it to add a (fake) live Stripe key to the env file, and it did it anyway. The hook only watched the Edit/Write tools, so Claude used a shell command. I called it out, Claude extended the hook to cover Bash, and in a new session the same request was blocked. To be fair, Claude had warned me when it first wrote the hook that it didn't cover shell commands. I just didn't read it.

Levels 6–7: clean context, then parallel agents

The fuller Claude's context window gets, the worse it does and the faster you burn usage. One person broke down their usage and found 95% of it came from conversations over 150,000 tokens. Clear your chat when you switch tasks, send big research jobs to subagents, and don't install every MCP server you see. Ponytail (about 173,000 stars) makes Claude reuse what exists and write less code, and it shows what each task costs. Its author claims about 45% fewer output tokens. JetBrains' independent test on 80 tasks found about 10% lower cost and about 15% less code. It's still worth using, but don't count on the big number.

Only once Claude can check its own work should you run more than one. I started two background agents, one adding an FAQ section and one adding testimonials. Then I opened agent view to watch both, each in its own git worktree so they don't step on each other. The FAQ agent worked 4:19 and the testimonials agent 3:23. Run one after the other, that's about 7:40. In parallel, both were done in about 5 minutes. I finished with one prompt: "Merge both branches, run the tests, and check it in the browser on desktop and phone." Boris's team still advises going slow: two or three lanes you can actually check beat 15 you can't. For more on this, see my Claude Code and AI agents pages.

Key takeaways

  • Keep CLAUDE.md short (under 200 lines) and tell Claude to update it after every mistake.
  • Give Claude a browser and a check, like tests, then let /goal loop until the check passes.
  • Put the rules that must never break in hooks, and test them like someone is trying to get around them.
  • Keep context clean: clear between tasks, use subagents, and be skeptical of big token-saving claims.
  • Run parallel agents in worktrees only after the basics are in place, and only as many as you can review.

FAQ

What is CLAUDE.md in Claude Code?

It's a file Claude Code reads at the start of every session so it remembers your project, commands, and past mistakes. You can generate it with /init, and Anthropic's docs recommend keeping it under 200 lines.

What does /goal do in Claude Code?

/goal keeps Claude working turn after turn until a check confirms your condition is true, such as "npm test passes with all three pricing tests green." In my demo, that took 41 seconds.

Can Claude Code get around a hook?

Yes, if the hook doesn't cover the tool it uses. My hook only watched Edit/Write, so Claude edited the .env file with a shell command. Once I extended the hook to Bash, Claude was blocked.

Does the Ponytail plugin really cut tokens by 45%?

That's what the author claims. JetBrains' independent test on 80 tasks found about 10% lower cost and about 15% less code.

Chapters

  1. 0:00Same tool, double the results
  2. 0:39Stop being the loop
  3. 1:27Level 1: Give Claude memory (CLAUDE.md)
  4. 2:26Join Shipping Skool
  5. 2:45Level 2: Give Claude eyes (Chrome DevTools MCP)
  6. 3:52Level 3: Skills (Impeccable)
  7. 4:46Level 4: /goal, /loop and auto mode
  8. 5:53Level 5: Guardrails and hooks
  9. 7:20Level 6: Keep context clean (Ponytail)
  10. 8:33Level 7: Parallel agents and worktrees
  11. 9:38Recap

Transcript

Read the full transcript

0:00Anthropic just looked at about 400,000 Claude code sessions. And here's what they found. People who were new to it got working, verified results 15% of the time. People who knew how to use it got about double that. Same tool, same model, same price. So the 100X isn't some secret model. It's how you use it. And in this video, I'm showing you exactly how the best builders use Claude code with real demos, not tips on a slide. We're giving Claude memory, giving it eyes, letting it check its own work, keeping it running without us, and running more than one Claude at a time. Let's jump right into it. So, here's the one idea this whole video is built on. Stop being the loop. Most people use

0:45Claude code like this. You ask, it writes code, you check it, you find the bug, you tell it, it fixes it, and you check again. You're the loop. You're the bottleneck. Boris Churnney, who leads Claude Code at Anthropic, put it like this. I don't prompt Claude anymore. My job is to write loops. And he's also said that if you give Claude a way to check its own work, it will 2 to 3x the quality of the final result. So, everything in this video is about one thing, getting you out of the loop safely. And I built a little demo project to show you all of it for real. It's a landing page for a habit app called Pulse. And like most AI made landing pages, it's pretty rough. Level one, give Claude memory. Every session,

1:31Claude starts from zero. It doesn't remember your project, your commands, or the mistake it made yesterday. That's what claude.md is for. It's a file Claude reads at the start of every single session. And you don't even have to write it. Type /init and you can see Claude reads the whole project and writes it for you. 24 seconds. It found the commands, how the project is laid out, and even a gotcha. This page needs a server to run because of how the scripts load. Now, here's the habit that compounds. Every time Claude makes a mistake, tell it. Update your cla.md so you don't make that mistake again. Over a few weeks, it stops repeating itself. But here's something to be wary about. Keep it short. Anthropic's own docs say under 200 lines because a bloated

2:16claw.md gets ignored. And if a rule has to happen every single time, that's not a cloud. MD thing, that's a hook. And we'll get to that real quick. Every video I make is built in code with Claude Code. Inside Shipping School Premium, I show you how stepby step to turn your product into videos that get views. Plus, new courses every day, 500 plus hours of replays, live calls every week, and a Q&A with me every Saturday. Join free. Link below. Level two, give Claude eyes. This is the big one. Claude can write a page, but it can't see the page, so it guesses unless you give it a browser. I added the Chrome DevTools MCP server, which is one command. Now watch.

3:01I just say, "Open the page in the browser, check the console, and fix anything that's broken." It opens a real browser, takes a screenshot, reads the console, and finds a script error that was firing on every single page load. It fixes it, reloads, and checks again. 48 seconds. And then it did something I didn't even ask for. It clicked the yearly pricing toggle and caught that my page was showing a price of $160, followed by about 14 zeros and a three. That's a feedback loop. The other MCP server the pros swear by is context 7. It feeds clog the current docs for whatever library you're using, so it stops writing code for an old version. And not to mention, if you want to learn Rea, the viral repo that lets Claude figure out how any app on your computer

3:47works, I just uploaded a detailed how-to in our school community. Link down below. Level three, skills. A skill is a folder of instructions that Claude only loads when it needs it. So, it costs almost nothing until it's used. And the one everybody's talking about right now is impeccable with about 85,000 stars on GitHub. It's a design skill that knows every tell of an AI made website. The purple tobble blue gradient, the emoji in a rounded square, cards inside of cards. I installed it as a plugin, which is two commands, and ran /impeccable audit. My page scored five out of 20. Critical. It flagged 14 AI template patterns, text that fails accessibility contrast, and a page that's almost twice as wide as a phone screen. Then

4:34/impeccable polish. About 3 minutes later, this is the before and this is the after. Same product, same sections, and it actually looks like a real company built it, which is insane. Level four, stop being the loop. Remember that broken price? There are three tests for the pricing, and all three are failing. So, instead of going back and forth, I type /goal npm test passes with all three pricing tests green. That's it. goal tells Claude to keep working turn after turn until a check confirms the condition is actually true. Not until it thinks it's done, until the tests say so. 41 seconds later, three out of three passing. This is exactly what Boris was

5:19talking about. You don't prompt every step. You set the condition and let it loop. There's also SL loop, which reruns a prompt on a timer while your session is open, like check my deploy every 5 minutes, and SLC schedule, which runs a task in the cloud on a schedule, even with your laptop closed. And here's one most people missed. Auto mode is now the default. Since late September, Claude Code starts in auto mode, where a separate safety model approves the safe actions for you instead of you clicking yes a 100 times. That's why these demos just run. But here's something to be wary about. Giving an agent more power only works if you set the rules. Right now, big US banks like JP Morgan and Morgan Stanley are finding AI agents

6:05inside customer accounts because people handed their bank login to AI assistants. To be clear, it's not a hack and nobody has reported losing money yet, but Morgan Stanley is blocking outside agents completely. So, how do you give Claude power safely? Hooks. A hook is a rule that runs every single time, no matter what the model decides. I asked Claude to add a hook that blocks it from reading or editing myv file, which is where API keys live. Done. Then I tested it. Add my livestripe key to the env file. It's a fake key, by the way. And it did it anyway because it didn't use the edit tool. It used a shell command, and my hook was only watching edits. So, I called it out and Claude extended the hook to cover shell

6:51commands, too. New session, same request, and this time it's blocked. And look at what it said. I didn't add the key. The project has a hook that blocks any command that touches and I'm not going to work around it. And to be fair, when Claude first wrote that hook, it actually warned me it didn't cover shell commands. I just didn't read it. So, that's the lesson. Read what your agent tells you. A guardrail only covers what you tell it to cover. So test your guardrails like someone's trying to get around them. Level six, keep Claude's context clean. Claude has a working memory called the context window, and the fuller it gets, the worse it does, and the faster you burn through your usage. A lot of people are complaining about usage limits right now. One of them broke down his own usage, and 95%

7:38of it came from giant conversations over 150,000 tokens. So clear your chat when you switch tasks. Send big research jobs to sub agents, which work in their own context, and just hand back a summary. And don't install every MCP server you see. Now, a plugin that's blowing up for saving tokens is Ponytail with about 173,000 stars. It makes Claude think like a lazy senior developer. Reuse what already exists and write less code. And it adds this bar above your prompt that tracks what each task is costing you. The author says it cuts about 45% of output tokens. JetBrains tested it independently on 80 tasks and found it was about 10% cheaper with about 15% less code. Still worth it, just don't

8:24believe the big number. Two other token savers did even worse in their tests. One that promises 60 to 90% savings actually cost more. Level seven now and only now multiply. Once Claude can check its own work, you can run more than one. Watch. Claude-BG, add an FAQ section. Claude-BG, add a testimonials section. Then, Claude agents, and you get this view. Two Clauds working at the same time, each in its own git work tree, so they never step on each other. And here are the real numbers because nobody shows these. The FAQ agent worked for 4 minutes and 19 seconds. The testimonials agent worked for 3 minutes and 23 seconds. One after the other, that's almost 8

9:10minutes. in parallel. Both were done in about five. So, no, it's not 100x. It's real time back and it stacks the more lanes you run. Then one more prompt. Merge both branches, run the tests, and check it in the browser on desktop and phone. And here's the final page. Boris calls work trees the single biggest productivity unlock, but his own team also says to slow down. Two or three lanes you can actually check beat 15 you can't. So, let's recap. Give Claude memory with a shortclaude.md. Give it eyes with a browser. Give it a check like tests and let slashgoal loop until they pass. Put the rules that must never break in hooks and test them. Keep its context clean and only then run more

9:56than one. That's how the same tool gets you double the results and it compounds every single week. Comment down below which one you're setting up first. And if this one helped you, hit the like button. It really helps more builders find it. Have a blessed day.

Ready to start building with AI?

Join the Shipping Skool community free. Paid members get live bootcamp calls every week and weekly-updated courses.

Join Shipping Skool

Keep reading