I’ll walk you through the latest methods I’ve created and currently use every time I work with Claude Code and Codex. Both are outstanding coding models that I rely on daily for programming. They perform impressively right from the start, but with the right approach, you can unlock even greater potential from them.
Here’s what this article will explore. I’ll share practical techniques you can start using immediately to get the most out of your coding agents. I highly recommend giving these a try as soon as possible, since I believe they can benefit nearly every developer.
I’ll introduce several precise techniques that range from straightforward prompt adjustments to broader mindset shifts you can adopt in your development workflow. Think of these as creative sparks to elevate your overall coding approach.
Why you should maximize Claude Code
Before diving in, I always like to explain why this topic matters. The reason to optimize your use of Claude Code and Codex is straightforward: with the right strategies, you can dramatically increase what these coding agents deliver for you.
There’s a popular saying that goes:
Those who benefit most from AI are already the most skilled
This means AI acts as a skill multiplier, not a leveler. It amplifies whatever abilities you already have.
Imagine programming skill as a point system where you earn points based on your coding ability and how efficiently you build solutions. For instance, if your current skill level sits at 10 points, AI might triple your effectiveness, bringing you to 30 points.
But if your baseline skill is already at 50 points, that same 3x multiplier pushes you to 150 points. The gap between skill levels was originally 40 points, and with AI, it widens to 120 points. This illustrates my core point: those who extract the most value from AI tools are the ones who were already strong. By applying targeted techniques to get more from your coding agent, you’ll see enormous productivity gains.
Specific techniques to maximize Claude Code
Now let me dive into the specific methods I use to get the most from Claude Code and Codex. I’ll cover four techniques in this article.
- Extensive use of OpenClaw and cron jobs. In short: use as many tokens as you can
- Leveraging Claude Code hooks effectively
- Ultracode pushes coding agents to work harder on complex tasks
- Have your coding agent display remaining tasks and a summary at the end of each response
Extensive use of OpenClaw
The first strategy I’ll discuss is making active use of OpenClaw and, broadly speaking, aiming to consume as many tokens as possible.
OpenClaw is a platform that lets you run bots within your messaging channels, such as Discord or Slack. These bots can be driven by an API connected to Claude Code or through your Codex subscription. The bots function as AI agents that operate either on a schedule (using cron jobs) or respond to particular messages or triggers. I find it easier to explain with concrete examples, so let me outline some specific scenarios where you can deploy OpenClaw agents.
- Set up an OpenClaw agent to notify you whenever you’re mentioned in a GitHub pull request and automatically conduct a code review
- Configure an agent to scan your product each night and deliver a report of issues by morning
- Deploy an agent to handle bug triaging automatically so you don’t have to sort through bugs manually
Naturally, there are countless other applications for OpenClaw. The core concept is that it’s essentially a coding agent operating around the clock, assigned to specific tasks, without requiring you to constantly steer the model. The model essentially makes all the decisions on its own.
Leveraging Claude Code hooks effectively
Claude Code Hooks is another powerful feature. Hooks are essentially scripts that execute at defined moments during your workflow. Some of the available hooks in Claude Code include:
- When Claude Code launches
- When Claude Code shuts down
- Whenever the agent poses a question to the user
- Whenever the agent completes a task
In essence, you can guarantee that a specific script runs whenever any of these events occur, along with several other trigger points. To give a concrete example, you could have Claude Code automatically extract and generalize insights from the current session when you close it. Or you could set it to play an audible alert on your computer whenever it asks you a question or wraps up a task.
The audible alert is something I set up recently, and it’s a technique I’m really pleased with. I configured Claude Code to trigger a sound on my computer whenever it asks me a question or finishes a task that needs my review.
This works wonderfully because I never need to keep my eyes on the terminal. I simply wait for the alert sound, and then I know it’s time to check on my agents. This makes it far easier to concentrate on other work once I’ve launched my agents.
Claude Code Ultracode – high-effort coding agents
Claude Code recently introduced a feature called Ultracode, which enables deeper, more intensive reasoning. This approach involves creating a swarm of agents that tackle a wide variety of tasks simultaneously, effectively a way to burn through more tokens while getting work done.
Generally speaking, using more tokens is beneficial because performance tends to improve with increased token usage. Of course, there’s a ceiling at some point, so spending a billion tokens on every
Time alone won’t lead to AGI, of course. Generally, the more tokens an agent uses, the better it performs—and Ultracode with Claude Code is designed to consume a large number of tokens.
You might wonder:
Does using more tokens mean the agent takes longer to finish a task?
The key insight here is that how long the agent takes isn’t the main concern. What truly matters is whether the agent completes the task correctly the first time, or whether you end up spending extra time fixing its errors.
To put it clearly, you’re essentially choosing between two approaches:
- Go with a faster, cheaper model that uses fewer tokens. It finishes the implementation in ten minutes, but then you spend an hour and a half correcting mistakes and iterating until you get the result you wanted.
- Spend 10 minutes briefing Claude Ultracode, then let it take 30 minutes to implement—40 minutes total. Yes, the upfront conversation takes more time, and the implementation phase is longer too, but you save significant time overall because you skip the lengthy back-and-forth of fixing errors and reworking the output.
When you frame it this way, the choice becomes obvious. You should almost always opt for the model that takes longer but delivers higher-quality results.
Show remaining tasks and provide a recap at the end of each response
Another helpful practice I’ve recently adopted is having my agents list any remaining tasks I need to handle, along with a brief recap at the end of each response. I set this up simply by adding instructions to my user-level CLAUDE.md file: at the end of every response, if you’re asking me to do something, always use the following format.
- []
- []
- []
... This makes it immediately obvious when the agent needs me to take action or run a test. I started doing this because I realized I couldn’t keep up with everything Claude Code outputs—it generates far too much text. Often, the agent would ask me to do something, but I’d skim right past it and miss the request entirely.
Naturally, this can hurt the quality of the work. But now that I use checkboxes, I spot pending tasks instantly and can address them without delay.
This is especially valuable when you’re juggling multiple agents at once and stepping away from a particular thread for more than 10 minutes. In those cases, it’s tough to remember exactly where you left off. You need a quick way to re-engage with the thread and know what to do next.
Along the same lines, I also have my coding agents include a recap right below the tasks they assign me. Claude Code already has a built-in recap feature, but I found it often appears too late. I prefer an instant summary, so I have Claude Code generate the recap itself. This makes it effortless to return to a thread after being away for 10 to 30 minutes working on other agents.
Conclusion
In this article, I shared the latest techniques I’m using to get the most out of Claude Code. I explained why continuously refining how you work with coding agents is so important—the people who excel are those who maximize the potential of these tools. If you want to stay ahead with the latest AI tools, you need to invest time in optimizing your workflow. I then walked through four specific techniques you can start using right away:
- Using OpenClaw
- Leveraging Claude Code hooks
- Maximizing token usage with Claude Code Ultracode
- Making sure Claude Code displays remaining tasks and recaps at the end of each response
I’m confident that applying these four techniques will lead to immediate improvements in your efficiency with Claude Code. I also encourage you to keep exploring new methods that can further enhance your workflow with coding agents like Claude Code and Codex.
Also check out my article on How to Run Multiple Coding Agents in Parallel.
👋 Get in Touch
👉 My free eBook and Webinar:
🚀 10x Your Engineering with LLMs (Free 3-Day Email Course)
📚 Get my free Vision Language Models ebook
💻 My webinar on Vision Language Models
👉 Find me on socials:
💌 Substack
🐦 X / Twitter
