Congratulations on your project! 🎉 I’ve build tools around RSS feeds before (they’re such a PITA!), so I know how much work this was, even with help from Claude!
Thank you for sharing your step-by-step process. 🙏 Very helpful!
A little bit of both, but I am definitely working on these tools to allow for better forecasting (and blog posts). For example, I'm directly using this application to write my next article and it's honestly turbocharged my effectiveness and ability to put otherwise disparate pieces together.
As for Claude Code, it's honestly a remarkable tool. I highly recommend just playing around with it. You'll find that once you get over the hurdle of 'I am not a programmer, it's scary' that you will find things you want computers to do that you can now make happen. I've written a dozen different, mostly little tools, but these were things that would have annoyed me or been manual toil a few years ago.
This is amazing and very inspirational. My fav bits from this:
1. "Have a design doc." Yes!
2. "Give your tools detailed instructions to avoid sadness. Explicit > implicit when working with AI. Over-communicate the steps. Describe exactly what you want, step-by-step, even if it seems painfully obvious." This is so key to get the best out of AI.
3. "Claude is a copilot, not an autopilot." Lol very true, always check the work the AI does it will get it wrong sometimes.
This post is inspiring to start my own project using Claude Code. Great work Carey!
Do it! Once you get over the mental block of your first, you'll be hooked. Also, let me know if you need a brainstorming buddy... or at least a human one.
Thanks Carey, I may just take you up on that at some point. I will need to have a careful think of what solutions I need for my own workflows and processes and where Claude Code could fit in that.
So glad you wrote this up. For someone who “doesn’t code”, this is a pretty cool tool :) I ended up adding the same x2 backoff on API call retries. A lot of commonalities here with the digest tool (which Karen Spinner originated) that we use for SheWritesAI. Our digest tool doesn’t feed articles to an AI API and is simpler in RSS handling thanks to more consistent RSS formats from just Substack, whereas you had more variability to deal with. If we ever open up our SheWritesAI digest beyond Substack, I’d need to extend it. I’ll check out your tool in detail soon!
Weirdly my comment didn't show up -- Is the SheWritesAI tool you mentioned available on Github? I'd love to compare (or maybe borrow -- I don't currently really handle Substack, but I should).
This is genuinely impressive work. The progression from simple RSS filtering to full AI-powered pipeline really captures the reality of building with LLMs right now, where the value isn't in th perfect upfront design but in iteratively solving actual friction points. The rate-limiting dance with the Claude API is spot-on, I hit that exact same wall when procesing batch data and had to learn the exponential backoff trick the hard way. What I apprecaite most is how honest this is about Claude's limitations, treating it like an overconfident intern is the exact right mental model. Most people either worship or dismiss AI tools, but this nails the practical middle ground.
Thank you. Yeah, I am perennially frustrated by the either/or mentality of so many folks. I'm glad you like the concept and my walkthrough -- do let me know if you have a chance to play with the tool. I really would love any feedback you have.
This is amazing! As a fellow info-junkie/avid-but-slow-and-distractible reader whose prone to shiny object syndrome and terrified by yet ready to battle rising autocracy I’m sooo tempted to try this out.
Just when I thought I was ready to finally commit to those other projects I’ve been procrastinating on….
Time to pet a cat while I reflect on those NY resolutions and priorities….
Oh you should -- I would love to see if you find value out of it, or what I could do to make it better / more useful. Let me know if there's any issues you run into.
Fun read, and super impressive too! I still need to think of a project to make use of all this cheap llm software engineering expertise available at my fingertips.
Couple of suggestions:
- when searching for "ai", you can massively reduce odds of false hits by searching instead for: "AI" (i.e. capital letters), " AI " (include spaces), "\nAI " (newline then AI), etc. But you'll want to include multiple options, e.g. "A.I." and "A.I" as well. Only searching for "artificial intelligence" has good odds of missing articles about AI.
- for unit testing. first llms should be able to do a good job of creating unit-tests based on your design doc. second, whenever you get a bug or unintended output, have habit of creating unit-test to check for that behaviour in the future.
So, for various reasons, including volume of keywords, I opted to make it case insensitive, so I honestly have no idea if I could make it case-sensitive for TLAs like AI.
I love the newline idea though... That's really clever and I hadn't thought of it. Also a good reminder re: unit tests on known bugs.
I'm glad you enjoyed it, and don't worry -- I started treating every minor little annoyance online as a potential AI software problem, and once you start with something (no matter how trivial), the ideas flow out of you.
Regular expressions (‘regex’) can flexibly handle things like optional periods in the AI acronym, mixed case, and requiring some form of whitespace or punctuation around it. Should be easy to write automated unit tests for ;)
Congratulations on your project! 🎉 I’ve build tools around RSS feeds before (they’re such a PITA!), so I know how much work this was, even with help from Claude!
Thank you for sharing your step-by-step process. 🙏 Very helpful!
Fascinating! Will you be using this process in your futures practice or was it intended more as a knowledge management challenge?
I keep reading about the brilliance about Claude Code and am slowly but surely getting FOMO.
A little bit of both, but I am definitely working on these tools to allow for better forecasting (and blog posts). For example, I'm directly using this application to write my next article and it's honestly turbocharged my effectiveness and ability to put otherwise disparate pieces together.
As for Claude Code, it's honestly a remarkable tool. I highly recommend just playing around with it. You'll find that once you get over the hurdle of 'I am not a programmer, it's scary' that you will find things you want computers to do that you can now make happen. I've written a dozen different, mostly little tools, but these were things that would have annoyed me or been manual toil a few years ago.
This is amazing and very inspirational. My fav bits from this:
1. "Have a design doc." Yes!
2. "Give your tools detailed instructions to avoid sadness. Explicit > implicit when working with AI. Over-communicate the steps. Describe exactly what you want, step-by-step, even if it seems painfully obvious." This is so key to get the best out of AI.
3. "Claude is a copilot, not an autopilot." Lol very true, always check the work the AI does it will get it wrong sometimes.
This post is inspiring to start my own project using Claude Code. Great work Carey!
Do it! Once you get over the mental block of your first, you'll be hooked. Also, let me know if you need a brainstorming buddy... or at least a human one.
Thanks Carey, I may just take you up on that at some point. I will need to have a careful think of what solutions I need for my own workflows and processes and where Claude Code could fit in that.
So glad you wrote this up. For someone who “doesn’t code”, this is a pretty cool tool :) I ended up adding the same x2 backoff on API call retries. A lot of commonalities here with the digest tool (which Karen Spinner originated) that we use for SheWritesAI. Our digest tool doesn’t feed articles to an AI API and is simpler in RSS handling thanks to more consistent RSS formats from just Substack, whereas you had more variability to deal with. If we ever open up our SheWritesAI digest beyond Substack, I’d need to extend it. I’ll check out your tool in detail soon!
Weirdly my comment didn't show up -- Is the SheWritesAI tool you mentioned available on Github? I'd love to compare (or maybe borrow -- I don't currently really handle Substack, but I should).
I'll DM you :)
This is genuinely impressive work. The progression from simple RSS filtering to full AI-powered pipeline really captures the reality of building with LLMs right now, where the value isn't in th perfect upfront design but in iteratively solving actual friction points. The rate-limiting dance with the Claude API is spot-on, I hit that exact same wall when procesing batch data and had to learn the exponential backoff trick the hard way. What I apprecaite most is how honest this is about Claude's limitations, treating it like an overconfident intern is the exact right mental model. Most people either worship or dismiss AI tools, but this nails the practical middle ground.
Thank you. Yeah, I am perennially frustrated by the either/or mentality of so many folks. I'm glad you like the concept and my walkthrough -- do let me know if you have a chance to play with the tool. I really would love any feedback you have.
Definitely saving this one to my RSS feed!
This is amazing! As a fellow info-junkie/avid-but-slow-and-distractible reader whose prone to shiny object syndrome and terrified by yet ready to battle rising autocracy I’m sooo tempted to try this out.
Just when I thought I was ready to finally commit to those other projects I’ve been procrastinating on….
Time to pet a cat while I reflect on those NY resolutions and priorities….
Oh you should -- I would love to see if you find value out of it, or what I could do to make it better / more useful. Let me know if there's any issues you run into.
Well done ! Thanks for sharing your knowledge.
Fun read, and super impressive too! I still need to think of a project to make use of all this cheap llm software engineering expertise available at my fingertips.
Couple of suggestions:
- when searching for "ai", you can massively reduce odds of false hits by searching instead for: "AI" (i.e. capital letters), " AI " (include spaces), "\nAI " (newline then AI), etc. But you'll want to include multiple options, e.g. "A.I." and "A.I" as well. Only searching for "artificial intelligence" has good odds of missing articles about AI.
- for unit testing. first llms should be able to do a good job of creating unit-tests based on your design doc. second, whenever you get a bug or unintended output, have habit of creating unit-test to check for that behaviour in the future.
So, for various reasons, including volume of keywords, I opted to make it case insensitive, so I honestly have no idea if I could make it case-sensitive for TLAs like AI.
I love the newline idea though... That's really clever and I hadn't thought of it. Also a good reminder re: unit tests on known bugs.
I'm glad you enjoyed it, and don't worry -- I started treating every minor little annoyance online as a potential AI software problem, and once you start with something (no matter how trivial), the ideas flow out of you.
Regular expressions (‘regex’) can flexibly handle things like optional periods in the AI acronym, mixed case, and requiring some form of whitespace or punctuation around it. Should be easy to write automated unit tests for ;)
Very cool what you’ve done here, Carey!
Oh yes. There's a little regex in the code in this section. I probably could improve the regex, but then I'd have two problems.
> pattern = r'\b' + re.escape(keyword_lower) + r'\b'
Please let me know! Is the SheWritesAI code available?