articles on strategies I apply when utilizing Claude Code to get probably the most out of it. Nevertheless, a subject I’ve spent much less time protecting is easy methods to enhance my Claude Code utilization on the whole. How do I optimize the way in which I work together with my Claude Code cases and the way in which my Claude Code operates throughout the code repositories I’m updating?
On this article, I need to spotlight how I’m frequently updating how I work together with Claude Code and the way my Claude Code operates, which makes me and my coding agent an increasing number of efficient over time.
The idea of continuous studying is extremely highly effective as a result of should you can enhance only a few % daily, the cumulative impact over weeks and months will grow to be huge. You may grow to be vastly extra environment friendly than simply the out-of-the-box model of Claude Code or every other coding agent.
Why carry out continuous studying?
I all the time attempt to cowl why a subject is vital, why you need to care about it, and the way it may help you. The explanation you need to carry out continuous studying is straightforward: should you’re simply utilizing the out-of-the-box model of Claude Code, Codex, or every other coding agent, you’re dropping out. In fact, these fashions are extremely highly effective, and should you evaluate them to only a few years in the past, you’ll be able to nonetheless be many instances extra environment friendly than beforehand.
Nevertheless, that doesn’t actually matter. What actually issues is that making use of continuous studying will once more offer you a large effectivity increase.
On this article, I’ll cowl one quite simple method on how I make my Claude Code enhance itself each single day, and likewise provide you with an perception into how I attempt to optimize my very own interactions with Claude Code to make the human coding agent interplay as efficient as doable.
Making Claude Code study from itself
I’ll begin off by protecting the straightforward method you could begin utilizing proper now, which is nearly definitely going to enhance how your Claude Code performs.
You may merely make a ability inside Claude Code that goes one thing like this:
Overview my final interactions with Claude Code from the final 24 hours.
Search for any issues that I encountered, issues that weren't working
effectively, and pointless instrument calling. Search for widespread errors
Claude Code was doing and different issues that may be optimized.
Look completely by way of all conversations and make a plan for a way we
can optimize our movement sooner or later, each inside every repository and
cross-repositories. Additionally search for insights that will be helpful for the
coding agent to know beforehand, each earlier than getting into a repository and
when working in a number of repositories on the similar time. Let’s say we name this ability review-past-performance. Now, what you need to do is about up a cron job to set off this ability at 2 am each night time or a while when you understand you’re not interacting along with your brokers actively.
What’s going to occur once you implement and run this ability is that Claude goes to undergo all of the previous conversations that you simply’ve had during the last 24 hours. It’s going to have a look at the threads. It’s going to see the place you bought caught with Claude Code (i.e., the place you spent extra time than you need to have), and it’s going to see the place Claude Code received caught making incorrect instrument calls, incorrect assumptions, or the place it merely didn’t have the context it wanted to carry out the duty successfully.
It’s then going to make a plan for easy methods to keep away from these items from occurring sooner or later and make Claude Code work extra successfully normally. It will implement adjustments equivalent to:
- Including extra info to brokers.md or related generic markdown information
- Creating particular expertise that the agent can both load on demand or run on demand when coping with sure duties
- Implementing particular scripts or tooling, equivalent to pre-commit hooks, testing scripts, and related, to keep away from errors from occurring once more
The very best half about establishing a cron job to run this ability every day is that you simply don’t even should work together with brokers in any respect. It’s going to have the ability to self-reflect in an environment friendly method, uncover inefficiencies, tweak them, and thus enhance Claude Code over time. Probably the greatest elements of that is that Claude Code goes to be personalized to your particular use circumstances. You might need a selected tech stack or preferences when working in repositories. Working this ability will uncover these preferences and optimize them to make them as environment friendly as doable.
By merely operating this cron job each night time, I’ve unlocked huge effectivity beneficial properties, the place my coding brokers have grow to be quite a bit stronger than they was, just because they make fewer errors. They’re extra conscious of the right method to doing issues, and general, they comply with my preferences higher.
Enhancing human interplay with coding brokers
One other extra sophisticated factor to optimize is the human interplay with coding brokers. I spend loads of time pondering and reflecting on easy methods to most successfully talk with my brokers to make them implement the code I would like as rapidly and as effectively as doable.
Clearly, this isn’t a solved downside but, as there are nonetheless loads of completely different tooling and platforms popping out to make coding brokers and interacting with them simpler, higher, and extra environment friendly. On this part, I’ll cowl a few of my reflections on the human interplay with coding brokers and the way I attempt to optimize it myself.
Notice that the strategies I’ll cowl are, in fact, optimized and tuned for my very own workflows, and I urge you to learn and find out about them and take into consideration how this is applicable to your individual workflows.
Working 7+ brokers directly
I usually discover myself operating loads of brokers directly just because I’ve loads of duties to finish and may begin engaged on them in parallel. In fact, there are exterior elements that determine whether or not it’s doable for me and even related to run this many brokers directly. When the scenario permits it and it is sensible efficiency-wise to take action, I’ll in fact run as many brokers as doable in parallel.
Nevertheless, I’ve discovered that once I begin reaching greater than seven brokers directly, I begin dropping management over all my brokers. I’m not capable of successfully context change between them, sustain with what every agent is doing, and successfully reply the agent when it’s asking me questions.
I’ve tried loads of completely different instruments and platforms to make this interplay extra environment friendly. I’m presently utilizing Warp, the place I take advantage of cut up panes for each tab once I’m working with parallel brokers inside one repository, and beginning new tabs for every completely different repository I’m engaged on. This works comparatively effectively, though, as talked about, I get caught when operating greater than seven brokers directly.
I’ve additionally tried extra IDE-based approaches like Conductor or Omnara, however I don’t really feel like they provide me any productiveness beneficial properties over what Warp can present me.
My takeaway from this part is a few strategies that will let you run as many brokers as doable directly. To begin with, the scenario has to permit it. It must be that you’ve sufficient duties that may all be run in parallel, and the place you’ll be able to let the agent run for lengthy sufficient that you simply’re not consistently interrupted. The 1st step is solely that the duty or duties that you simply’re finishing have to permit it.
Second of all, a really highly effective factor when working with many brokers in parallel is a recap. Claude Code has began offering a recap on the backside of the chat, which is extremely highly effective. It offers you an excellent transient overview of what you’re doing in that chat, which lets you rapidly atone for context when you need to work together with an agent once more. I urge you to allow recaps and actively use them if you must learn up on the context of a selected thread.
Lastly, on this part, I might additionally like to notice that Claude Code, at present, as of the writing of this text, simply launched an agent view in Claude Code. It is a view that ought to make it simpler to maintain an summary of all of your brokers directly. I haven’t tried it myself but, although it seems to be to take care of precisely the issue I’m describing on this part. I’ll positively be making an attempt it out and writing an article on it sooner or later.
Let the agent ask you questions, not the opposite method round
This subsection is an fascinating one as a result of the widespread option to work together with AI fashions, not less than at first, was to ask them questions and have them reply to you in a concise method. Nevertheless, this utterly shifts when you begin coping with long-running code classes. You don’t need to ask it questions anymore, you need it to work as independently as doable for so long as doable, and solely cease when it has to ask you questions.
That is thus one thing I like to recommend that you simply enter into the prompts of your coding brokers. You need them to run for as lengthy and independently as doable and solely cease implementing as soon as it has to ask the consumer a query. This, in fact, additionally ties strongly into one other article I’ve written, which is easy methods to let Claude Code validate its personal work. To make the agent run for a very long time, you must give it an choice or risk of verifying its personal work, which I coated in one other In the direction of Information Science article. Test it out under:
How to Make Claude Code Validate its own Work
Conclusion
On this article, I coated how I frequently enhance my Claude Code occasion, each by making Claude Code enhance itself by way of self-reflection each night time and thru enhancing human interplay with Claude Code and different coding brokers. I consider each of these items are issues that you need to attempt to optimize as an engineer to make your coding simpler. As an engineer, you need to all the time be trying on the subsequent bottleneck: what’s slowing you down probably the most and would unlock the best productiveness increase. For me, I found that this was:
- Claude Code repeating errors, that are fastened by the primary part on this article
- The human interplay with Claude Code, which I coated within the second part of this text
I urge you to consistently be in search of such bottlenecks and attempt to take away them as rapidly as doable to make your coding efforts as productive as they are often.
👉 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
👉 Discover me on socials:
💌 Substack
