I use a conventions file. But I have a conventions file for many different types of tasks. Then I compose them for the specific thing I'm doing. I'm thinking of making a conventions repo to house them. Then I use the /ask mode to rubber duck out a design, then I have it help me write an implementation plan with coding examples. I do this part when I have many of the same tasks to do. At the bottom of this file, I have it provide implementation steps where I have it specify exactly what to do and when to stop and not go into the next step. Let's say there are 7 total steps where steps 3-4 are somewhat complex that need to be done 0-n number of times per task. Then I use scripted aider to loop over all the source files or whatever and for each file I do a aider run telling it to read the implementation plan, add any specific conventions plans that are applicable to that step if applies, then my run command is to do step #1 in the plan doc. Then I do the same for step 2. Then steps 3-4 need to happen for each subsection lets say. For for every subsection in the file, I say do steps 3 and then step 4. Then back out of the nested loop I tell it to do the steps 5-7 one at a time. I can basically use a single prompt and subprompt it easily to get the best out of the ease of a single prompt but also the accuracy of a multi-prompt. I need to write a conventions file for writing conventions files and these implementations plans and these scripted aider tasks. But I have aider write them all right now anyway, it would just go faster.
Yeah this is a gold mine waiting to be unlocked for many AI powered Engineers. You've essentially built your own ai coding agent on top of Aider that very likely outperforms on your codebases. More content to come on this topic. It's going to break the eng world honestly once ppl realize there are basically no limits on the type of ai coding pipelines you can build on top of tools like Aider.
@@indydevdan your next tutorial you should show the kind-cmd to do on save actions like formatting with black and isort. I left a different comment on this video with instructions. It’s a HUGE boost.
this was great but very fast to absorb as a newbie, would be good to see more of each step and stages of what was happening, but thank you for showing :)
This stuff has been an amplifier for me. By trade i'm not a developer or engineer, i'm a sales guy lmao. I've started by making an agent to tutor me on coding and using Kali Linux tools. Two years ago, my mind wasn't even on this subject, all changed overnight.
Keep it up man! 🙌🏾 I transitioned from sales into data over the last 3 years. Was hard but for sure best decision ever. It's an unfair advantage to both sell and build.
3:00 I declare the types where they are needed and make the most sense in the context. A additional file for the types is an unneccessary, additional file.
Your content is awesome. Besides the AI stuff, I found little gems like those senv, mkenv, okey commands very useful. Probably the best invested time I can remember watching TH-cam videos. I'd be very interested on a AI driven development course from you.
Thanks, nice video. Coding's not my full time job, but I still do it from time to time for small stuff. I've always avoided test case writing beyond simple printing of debug info as it seemed like more work, but you're selling me on the utility even in my "if it works at all, it's good enough" use cases. If I can lean on AI more heavily and reliably by way of testing since that grounds the AI, then it'd definitely be worth the time to do things (more) right.
Thank you dude, that's exactly what I was looking for. Love your videos. Don't have time myself, but with your help, I'll probably be able to switch after all. Currently using only Cursor with its compose. Although i don't like that he is sometimes making mistakes and i spend more time checking those. Guess need to use your idea with tests.
Always inspiring to watch you doing stuff the AI-way. Love the way you make cursor and aider go hand in hand on that. I am wondering if you tried using Cursors Composer instead of aider. And if so what is your take on these compared.
Definitely have see previous video on channel. Cursor's composer is a big deal because it represents 'mainstream' adoption of ai coding in the tech ecosystem. My take is: Aider >= Cursor BUT (Aider + Cursor) > Aider
look forward to some more videos from ya. id love to see more of the notion wrapper too...also what are using to get the transparent background on everything
This was really great Dan. Will your AI coding class be suitable for beginners or if not where would you point a highly motivated beginning learner to get the basics? Thanks again
Great video! Do you keep the context in your aider for the whole coding session or do you restart aider? Does it increase the API costs much to keep the context? Also, if you edit first using aider, then using Cursor, and then again with aider, does aider consider the latest updates from cursor?
From the video it seems he doesn't reset the context. Using the claude --cache-prompts option should lower the API costs. if on, aider includes the following in cache: system message, read only files and repository map. The repo map might be changed (and cache reset) if files are added or deleted, based on --map-refresh option. If I use the logs --llm-history-file, it seems that aider keeps the context for the further prompts reasonably small, it probably doesn't include the full session history. If in doubt, one can run /clear to clear the chat history. And for the 2nd question, aider always sends an up to date version of the file with each prompt.
Great Q. I focus on one 'feature' from a UX (not DX) perspective at a time. Generally that is the largest a single coding session will become: the length of the coding session required to ship that feature. If I find the LLM/Aider becoming confused about any of the context, I '/clear' it and start fresh. To be clear though, I prioritize 'time to feature completion' and not cost of context window. The API costs do increase for the larger features but I recommend ignoring them and eating the cost to get the practice in pushing these tools to the limits of their efficiencies. LLM prices are going to zero so it's best to plan/code/practice/build for that future.
Is it possible to do this without Cursor? I think I saw a video of vscode plugins for aideer, continue, and cluadedev. I'm new to all of this so I appreciate any pointers.
@indydevdan another awesome video mate. can you suggest (or show?!) how I can use this setup for google apps (or google ads) scripts?? I'm building more & more complicated files now, but don't have any tests (all single page scripts, some ~5000 LOC) all a bit manual currently (copying code from cursor to sheets/ads to run on live data) I get I can have mock data & test against that... but this video has opened my eyes.. I ralise I can automate this way more. any pointers?
I use only cursor, while making a ts game in cocos creator. What's the added benefit of aider, still not sure after watching this demo 😮 the multi file editing or what? (I move from file to file in cursor and just select all text and run the ctl-k edits)
Aider has several additional benefits such as running tests post code generation, auto lint fixes, conventional commits, larger context support (cursor is limited to 10K or so tokens), etc. Aider is definitely trickier to use though because it's just CLI.. and I'm still trying to rapidly stall my developing Composer addiciton. :)
I agree with @yurijmikhassiak7342. Needs more time, but these 'ai software engineers' like the Replit agent, devin, and microsofts version (will likely merge with devin) are the next generation of ai tools. We will cover them as they emerge. Stay locked in.
Thank you! I'm playing with this idea right now. It will likely ship after the ai coding course and will be a high ticket, highly opinionated, information dense membership for those who want to gain a bleeding edge take on LLMs, ai coding, and agentic engineering as a whole. It's a great place to share things I honestly can't justify sharing anywhere at zero cost. We'll see how things develop.
I think you're not using Aider/Cursor to the best of their abilities. I've been experimenting a lot and I find you should really be trying more and more to tell these tools _what_ you want your _application_ to do, not what you want the code to be thus moving further up the abstraction chain. Quite often here you're giving it commands such as "update x.ts: add some method and a test for it". Why not write a more verbose prompt telling it what you're actually trying to accomplish on a higher level and let it work out how to do it? Personally I've found it to produce better solutions than I thought of myself especially with more complex problems and even if it doesn't get it right, I can always iterate with it and we will converge on a better solution. Just a thought but in general love the content. I'm fully invested in these tools now after watching your videos so much appreciated!
No, you are right I can be pushing Aider and Cursor further here. There are different AI Coding prompt 'abstraction levels' and focusing on 'what' over the 'how' is a higher level. It can do more, but also is more error prone in larger production codebases (for now). In this video you're seeing concise, lower-level, more descriptive AI code prompts. 🙏 Glad to here you are getting value out of these tools + videos.
How about using .cursorrules as a read-only file in Aider. This way, we will be able to use both Cursor and Aider using the same coding conventions.
This is the most informative and useful AI coding video I've seen so far. Thank you for this.
I use a conventions file. But I have a conventions file for many different types of tasks. Then I compose them for the specific thing I'm doing. I'm thinking of making a conventions repo to house them.
Then I use the /ask mode to rubber duck out a design, then I have it help me write an implementation plan with coding examples. I do this part when I have many of the same tasks to do. At the bottom of this file, I have it provide implementation steps where I have it specify exactly what to do and when to stop and not go into the next step.
Let's say there are 7 total steps where steps 3-4 are somewhat complex that need to be done 0-n number of times per task.
Then I use scripted aider to loop over all the source files or whatever and for each file I do a aider run telling it to read the implementation plan, add any specific conventions plans that are applicable to that step if applies, then my run command is to do step #1 in the plan doc. Then I do the same for step 2. Then steps 3-4 need to happen for each subsection lets say. For for every subsection in the file, I say do steps 3 and then step 4. Then back out of the nested loop I tell it to do the steps 5-7 one at a time.
I can basically use a single prompt and subprompt it easily to get the best out of the ease of a single prompt but also the accuracy of a multi-prompt.
I need to write a conventions file for writing conventions files and these implementations plans and these scripted aider tasks. But I have aider write them all right now anyway, it would just go faster.
Yeah this is a gold mine waiting to be unlocked for many AI powered Engineers. You've essentially built your own ai coding agent on top of Aider that very likely outperforms on your codebases. More content to come on this topic.
It's going to break the eng world honestly once ppl realize there are basically no limits on the type of ai coding pipelines you can build on top of tools like Aider.
@@indydevdan I seriously owe so much to your videos! Thanks!
@@indydevdan your next tutorial you should show the kind-cmd to do on save actions like formatting with black and isort. I left a different comment on this video with instructions. It’s a HUGE boost.
this was great but very fast to absorb as a newbie, would be good to see more of each step and stages of what was happening, but thank you for showing :)
This stuff has been an amplifier for me.
By trade i'm not a developer or engineer, i'm a sales guy lmao.
I've started by making an agent to tutor me on coding
and using Kali Linux tools.
Two years ago, my mind wasn't even on this subject,
all changed overnight.
Cool. Sounds very interesting. Would love to hear more about it.
Keep it up man! 🙌🏾
I transitioned from sales into data over the last 3 years. Was hard but for sure best decision ever. It's an unfair advantage to both sell and build.
I am also in sales & have even gotten paid to dev an application before cursor existed. Now when I'm not in an appt, I'm building software with AI.
Awesome job Dan! Keep up the high quality of Your vids! Tks
I went and did a bunch of this and I'm already seeing an improvement
Cursor has been soooo fking fun to play with. Been making chrome extension, even got it to make an AI based tamagachi pet.
What's a tamagachi pet?
3:00 I declare the types where they are needed and make the most sense in the context. A additional file for the types is an unneccessary, additional file.
Your content is awesome. Besides the AI stuff, I found little gems like those senv, mkenv, okey commands very useful. Probably the best invested time I can remember watching TH-cam videos. I'd be very interested on a AI driven development course from you.
I'm pretty new to coding and I'm def a little lost in this video but I'll have to come back as I learn more!
Thanks, nice video. Coding's not my full time job, but I still do it from time to time for small stuff. I've always avoided test case writing beyond simple printing of debug info as it seemed like more work, but you're selling me on the utility even in my "if it works at all, it's good enough" use cases. If I can lean on AI more heavily and reliably by way of testing since that grounds the AI, then it'd definitely be worth the time to do things (more) right.
Thank you dude, that's exactly what I was looking for. Love your videos. Don't have time myself, but with your help, I'll probably be able to switch after all. Currently using only Cursor with its compose. Although i don't like that he is sometimes making mistakes and i spend more time checking those. Guess need to use your idea with tests.
Always inspiring to watch you doing stuff the AI-way. Love the way you make cursor and aider go hand in hand on that.
I am wondering if you tried using Cursors Composer instead of aider. And if so what is your take on these compared.
Definitely have see previous video on channel. Cursor's composer is a big deal because it represents 'mainstream' adoption of ai coding in the tech ecosystem.
My take is: Aider >= Cursor
BUT
(Aider + Cursor) > Aider
Which theme are you using for VSCode/Cursor? It looks so fancy 🙂
Really thanks!. Keep doing this!. Pioneering software development.
look forward to some more videos from ya. id love to see more of the notion wrapper too...also what are using to get the transparent background on everything
this is excellent
This was really great Dan. Will your AI coding class be suitable for beginners or if not where would you point a highly motivated beginning learner to get the basics? Thanks again
If you can code you will be able to learn to code with AI through this course ESPECIALLY if you are a highly motivated beginner.
Man, you are amazing!
I'm a beginner coder, where can I learn all the little tricks and keyboard shortcuts you use to do things so quickly in cursor?
Great video! Do you keep the context in your aider for the whole coding session or do you restart aider? Does it increase the API costs much to keep the context?
Also, if you edit first using aider, then using Cursor, and then again with aider, does aider consider the latest updates from cursor?
Was wondering the same
From the video it seems he doesn't reset the context.
Using the claude --cache-prompts option should lower the API costs. if on, aider includes the following in cache: system message, read only files and repository map.
The repo map might be changed (and cache reset) if files are added or deleted, based on --map-refresh option.
If I use the logs --llm-history-file, it seems that aider keeps the context for the further prompts reasonably small, it probably doesn't include the full session history.
If in doubt, one can run /clear to clear the chat history.
And for the 2nd question, aider always sends an up to date version of the file with each prompt.
Great Q. I focus on one 'feature' from a UX (not DX) perspective at a time. Generally that is the largest a single coding session will become: the length of the coding session required to ship that feature.
If I find the LLM/Aider becoming confused about any of the context, I '/clear' it and start fresh.
To be clear though, I prioritize 'time to feature completion' and not cost of context window. The API costs do increase for the larger features but I recommend ignoring them and eating the cost to get the practice in pushing these tools to the limits of their efficiencies.
LLM prices are going to zero so it's best to plan/code/practice/build for that future.
@@indydevdan thanks for the response!
Great video man. Thanks. Have you tried cline (Claude dev) ?
So with Cursor Composer, would there be a benefit in using Aider as well?
Is it possible to do this without Cursor? I think I saw a video of vscode plugins for aideer, continue, and cluadedev. I'm new to all of this so I appreciate any pointers.
@indydevdan another awesome video mate. can you suggest (or show?!) how I can use this setup for google apps (or google ads) scripts??
I'm building more & more complicated files now, but don't have any tests (all single page scripts, some ~5000 LOC)
all a bit manual currently (copying code from cursor to sheets/ads to run on live data)
I get I can have mock data & test against that... but this video has opened my eyes.. I ralise I can automate this way more. any pointers?
I use only cursor, while making a ts game in cocos creator. What's the added benefit of aider, still not sure after watching this demo 😮 the multi file editing or what? (I move from file to file in cursor and just select all text and run the ctl-k edits)
Aider has several additional benefits such as running tests post code generation, auto lint fixes, conventional commits, larger context support (cursor is limited to 10K or so tokens), etc. Aider is definitely trickier to use though because it's just CLI.. and I'm still trying to rapidly stall my developing Composer addiciton. :)
di ZED next, please. i love Zed so much it works on Linux without fail... such a blessing for Linux user zed is
Aider appears to "know" the notion API without any extra configuration. How come?
What is your thought on Replit AI agent ?
Have you used codeium as it has more of a tdd focus ?
I wonder if I can get Pythagora to write such a wrapper library on its own, with even less prompting.
Do you run into token limits with Claude while using Aider/Cursor?
what are you cooking? :)
What do you think of Replit Agent?
slow and crushing, it's not a working product as for now. I have canceled my subscription.
@@yurijmikhassiak7342 yeah it fails on anythign remotely complex - and it attempts to do everything too so yeah
I agree with @yurijmikhassiak7342. Needs more time, but these 'ai software engineers' like the Replit agent, devin, and microsofts version (will likely merge with devin) are the next generation of ai tools. We will cover them as they emerge. Stay locked in.
waiting for the OpenAi o1 update from you
That's useful will put it in a .cursorrules file
Great job! Why don't you make a membership thing on youtube. I'd do $10/mo for this stuff.
Thank you! I'm playing with this idea right now. It will likely ship after the ai coding course and will be a high ticket, highly opinionated, information dense membership for those who want to gain a bleeding edge take on LLMs, ai coding, and agentic engineering as a whole. It's a great place to share things I honestly can't justify sharing anywhere at zero cost. We'll see how things develop.
Whats up Daniel?
lesgooooo
Deepseek v3 rules!
Anybody else use these for writing IaC code? It's.. interesting
I like to store my typescript files however my AI likes to store them.
I think you're not using Aider/Cursor to the best of their abilities. I've been experimenting a lot and I find you should really be trying more and more to tell these tools _what_ you want your _application_ to do, not what you want the code to be thus moving further up the abstraction chain. Quite often here you're giving it commands such as "update x.ts: add some method and a test for it". Why not write a more verbose prompt telling it what you're actually trying to accomplish on a higher level and let it work out how to do it? Personally I've found it to produce better solutions than I thought of myself especially with more complex problems and even if it doesn't get it right, I can always iterate with it and we will converge on a better solution.
Just a thought but in general love the content. I'm fully invested in these tools now after watching your videos so much appreciated!
No, you are right I can be pushing Aider and Cursor further here. There are different AI Coding prompt 'abstraction levels' and focusing on 'what' over the 'how' is a higher level. It can do more, but also is more error prone in larger production codebases (for now). In this video you're seeing concise, lower-level, more descriptive AI code prompts.
🙏 Glad to here you are getting value out of these tools + videos.
I do not understand. Can you make it more beginner friendly
Too small, I watch on phone at night 😢
ACTION PACKED FIRST
👀