There is always more work than time. These protocols help you choose what matters, protect focused engineering work, and finish the day ready for tomorrow.
I agree to start the day with plan and have some dedicated time to make the change.
Essential point 3, 4, 5 need to be merge into one single pointers.
Once you decided to must make a change, observe and understand the change, chunk it out and take the smallest piece to solve the problem, code and solve that smallest piece with the testing ie close the feedback.
Point 3 is essentially about not asking AI to do something, but asking AI to explain things and you draw your own conclusion.
So there's a certain difference between asking AI, add a button to this page or asking AI and checking with the code, with the test, with documentation that is around there, how, what is the architecture that the application is following and then based on that we decide how and where to add the button.
This example of a button may seem very simple, but for other more complex examples, especially when we don't understand the output the AI generated and why I generated it in this way, that moment is when we are failing to do this point 3.
I think what defines a protected, uninterrupted deep work block has changed substantially. For example, you write that you disable all notifications. I don't, especially from harnesses - cmux, claude code, etc.
With the agentic loop, I cannot really do that anymore because I don't have agents working on separate things in parallel.
One might be reviewing PRs, the other might be asking for a follow-up question on a grilling session. Once I answer that, I immediately jump to the other agent waiting for my input
So, I think this used to work because we didn't programme with agents, but with agents, what counts as an interruption has completely changed.
Yes, I also have my notifications disabled for people but not disabled for my agents 😅 😅 . I'm using herdr.
I think the point is that all your work sessions of agents are about work. And yeah, you may be working in different things, but I think it's the equivalent of working in different code files in the past. You would be jumping from one file to another as you reference them and check them instead of just working in doing one file until you finish it.
That's more a waterfall than deep work.
I think the anti-pattern would be that you are working with your agents but some agents are about coding, one agent is doing grocery list for shopping, and then another agent is talking on a conversation about your plans for the weekend. That can easily remove your focus and make you forget what were you doing in the first place with the code.
But yes, I am also working in multiple implementation tasks at the same time with agents. And I think instead of thinking AI has to be synchronous and wait for the response, it's better to be having multiple loops or workflows happening in parallel. Also many of them in the background that are triggered automatically and I only arrive to the result.
These are good 10 pointers.
But I see 10 pointers are too heavy as well.
I agree to start the day with plan and have some dedicated time to make the change.
Essential point 3, 4, 5 need to be merge into one single pointers.
Once you decided to must make a change, observe and understand the change, chunk it out and take the smallest piece to solve the problem, code and solve that smallest piece with the testing ie close the feedback.
Yes, they have to be done together!
Thank you! Can you give some examples of point 3. reading code part or best practices point. I want to work on this.
Point 3 is essentially about not asking AI to do something, but asking AI to explain things and you draw your own conclusion.
So there's a certain difference between asking AI, add a button to this page or asking AI and checking with the code, with the test, with documentation that is around there, how, what is the architecture that the application is following and then based on that we decide how and where to add the button.
This example of a button may seem very simple, but for other more complex examples, especially when we don't understand the output the AI generated and why I generated it in this way, that moment is when we are failing to do this point 3.
I think what defines a protected, uninterrupted deep work block has changed substantially. For example, you write that you disable all notifications. I don't, especially from harnesses - cmux, claude code, etc.
With the agentic loop, I cannot really do that anymore because I don't have agents working on separate things in parallel.
One might be reviewing PRs, the other might be asking for a follow-up question on a grilling session. Once I answer that, I immediately jump to the other agent waiting for my input
So, I think this used to work because we didn't programme with agents, but with agents, what counts as an interruption has completely changed.
Yes, I also have my notifications disabled for people but not disabled for my agents 😅 😅 . I'm using herdr.
I think the point is that all your work sessions of agents are about work. And yeah, you may be working in different things, but I think it's the equivalent of working in different code files in the past. You would be jumping from one file to another as you reference them and check them instead of just working in doing one file until you finish it.
That's more a waterfall than deep work.
I think the anti-pattern would be that you are working with your agents but some agents are about coding, one agent is doing grocery list for shopping, and then another agent is talking on a conversation about your plans for the weekend. That can easily remove your focus and make you forget what were you doing in the first place with the code.
But yes, I am also working in multiple implementation tasks at the same time with agents. And I think instead of thinking AI has to be synchronous and wait for the response, it's better to be having multiple loops or workflows happening in parallel. Also many of them in the background that are triggered automatically and I only arrive to the result.