I've been maintaining a personal calorie tracker (keeps me entering), and using claude for entering meals has been the perfect experience.
One of my favorite features has been giving a recipe url and having it figure out the calories for that recipe with line-by-line entries for each ingredient. Then it keeps the extracted recipe in markdown format for the next time I want to make it.
Substitutions are also great - "I'm making [such and such recipe], but with half the sugar."
One of the others has been "I had a meal a couple weeks ago at [x restaurant], can you enter that again for right now?" Having the long-term context for entry has been excellent. And a related one, where I upload a photo of the menu, and say which menu items I'm eating and it puts together an estimate for the calorie count.
I lean heavily on a long auto-compacting claude code session for all of this. I have a UI for graphs, history, and simple entries, but claude's been my favorite means of entry by far
I also appreciate the social break my brain seems to need in order to keep socializing. I'm an introvert but also social, so I love to talk to friends and hang out - but I still need that 5 minutes to just stare off into space and recharge before I jump back into the fray.
I quit smoking about 15 years ago, but still vape - and I remain a huge fan of the easy excuse to step away for 3-5 minutes and just zone out, or change from a large group to a smaller 2-person conversation outside.
Even back when we could smoke in bars, I would still sneak outside from time to time to settle and recharge.
Security questions have always been ridiculous, but I'll especially never understand how "favorite [thing]" ever made it to production anywhere. "Favorite movie" can change multiple times in the same conversation.
This is so well done and very cool. Thanks for building it and offering it.
As someone who hasn't had a piano lesson in about 40 years, I find myself wanting to play with the keys to match the melody. So I hear the initial melody, and then I'm practically hitting keys at random (guessing where I should be on the keyboard) until I find the first note, and then I have to listen over and over again while trying to find the second note. I kind of want to hunt and peck until I'm ready and then get tested to see if I nailed it.
I've done a similar thing with close friends and family who would constantly ask me things I couldn't possibly know because I always came up with an answer.
Eventually I realized why and explained, "you know, I'm really just going to do a web search for what you just asked me, and maybe a couple more until I have a decent answer and then give you that answer. Let me show you how I would go about that".
From then on, they started getting into the habit of doing that for themselves. I think now with LLMs, they've kept the habit, but the LLM gives a more complete answer with fewer steps so it becomes the default. I think the magic of AI is two-fold (well, more than two, but two bullets for this conversation).
1. You don't have to "query". You can just braindump, and it'll build a context and figure out what you're looking for
2. It's conversational, so instead of filtering and tweaking results from the first query, your second "query" builds on top of the context from the first question, and you get a stronger result as the conversation continues.
I've had cases where it doesn't explicitly use a skill I've added BUT it still performs the actions described in the skill on its own more often than it did before I created the skill. I'd rather it use the skill for consistency, but having it follow most of the steps most of the time in cases I've forgotten to explicitly call out the skill is a better outcome.
I really appreciate this post. It reminds me of a video I watched a couple years ago that does an excellent job of demonstrating how culling works with actual code and visuals
I'm similarly unpredictable in my home. Add to that the others in my house, and it's impossible to even guess what everyone's intentions are at any given time.
Sometimes I daydream about a "solo mode" where the timings on lights are tighter and my music can follow me around the house when I'm up at night and nobody else is. But most times I'm trying to find the get-out-of-way averages that keep everyone happy.
Some things work great: Automated lights everywhere. Automated dimming of lights at night or sunset or whatever. Notifications when the laundry is done, or the cat litter is ready to be changed, or someone is at the door, or the garage door has been left open - all great. What music to play in what room at any time? Always changes. When to "dim all the lights" because Plex started a movie? But my son is building Legos in the dining room, and my wife is knitting and needs the couch light on. Sometimes I want it, but not every time.
For those things having a single button press is still a huge win over opening multiple apps and getting the right things set the right way for each participant.
I feel like dealing with robo-calls for the past couple years had led me to this conclusion a bit before this boom in ai-generated text. When I answer my phone, if I hear a recording or a bot of some sorts, I hang up immediately with the thought "if it were important, a human would have called". I've adjusted this slightly for my kid's school's automated notifications, but otherwise, I don't have the time to listen to robots.
I have the same issue. Even when I ask it to do code-reviews and very explicitly tell it not to change files, it will occasionally just start "fixing" things.
My house is 150 years old, but it's also been rebuilt several times in that time. My neighbors' houses are less than a decade old. We have all swapped stories about replacing things behind drywall. Leaks. Electrical issues. Ducts. Everything. Consider yourself lucky if you have not. 50 years is not the number you should bet upon.
In Chicago, code requires EMT for all electrical, which can be annoying for adding a new run, but at the very least it makes it less likely for rodents to chew through or other interference.
After wiring my whole house with Ethernet and ceiling speakers, and now dealing with a couple leaky pipes and several problems from previous owners, I'm considering ways to make these things easily accessible/replaceable while keeping an eye toward aesthetics.
For sure. I've wired my old house with speakers in every ceiling, and cat-6 in every room. I've had a small pipe burst and a couple leaks behind a bathroom.
I've patched quite a bit of drywall, and I'm about mediocre at it. But it seems so silly and unnecessary to me.
Everything else in this world that requires maintenance comes with access panels and other means of easy access. In our living spaces, some of which should ideally last tens of years (mine is from the 1890s), we seal it all away.
I watched a video recently, which I can't find, where an architect set up a beautiful wooden baseboard around the entirety of their property, and that baseboard held all mechanicals and was perfectly clean and easy to get into as needed.
Drywall is manageable and cheap, I agree. But it's more painful than it should be for something that _will_ require maintenance.
I will never understand why we fill our walls with mechanical and electrical infrastructure and then wrap them in a paper and plaster, which then needs to be torn, broken, and repaired in order to maintain said infrastructure.
Pipes will fail. Wires will fail. Ducts will fail. Maybe not in 5 years, but over the span of 20, they will. Why make them so frustratingly inaccessible?
I can only wear tall-size clothing, and generally I've found that none of my t-shirts shrink "in", but they _all_ shrink "up". I can make them last longer washing them delicate and "air-drying" (in the dryer, light or no heat), but eventually they all get shorter. I have to replace most of my undershirts annually, and I rarely bother with t-shirts anymore.
I haven't used cursor, so I'm not sure I can be much help there. I've been mostly using claude code and IntelliJ IDEs for code-reviews when necessary. Over the past year I've moved to almost entirely coding via agent. Maybe my input will be helpful.
One very important thing to keep in mind is context management. Every time your agent reads a file, searches documentation, answers a question, writes a file, or otherwise iterates on a problem, the context will grow. The larger the context gets, the dumber the responses. It will basically start forgetting earlier parts of the conversation. To be explicit about this, I've disabled "auto-compact" in claude code and when I see a warning that it's getting too big, I cut things off, maybe ask the agent to commit, or write a summary, and then /compact or /clear. It's important to figure out the context limits of the model you're using and stay comfortably within them.
Next, I generally treat the agent like a mid-level engineer who answers to me. That is to say, I do not try to convince it to code like I do, instead I treat it like a member on my team. When I'm on a team, we stick to standards and use tools like prettier, etc to keep the code in shape. My personal preferences go out the window, unless there's solid technical reason for others to follow them.
With that out of the way, the general loop is to plan with the agent, spec the work to be done, let the agent do the work, review, and repeat. To start, I converse with the agent directly. I'm not writing a spec, I'm discussing the problem with the agent and asking the agent to write the spec. We review, and discuss, and once our decisions are aligned and documented, I'll ask it to break down how it would implement the plan we've agreed upon.
From there I'll keep the context size in mind. If implementation is a multi-hour endeavor, I'll work with the agent to break down the problem into pieces that should ideally fit into the context window. Otherwise, by this point the agent will have asked me "would you like me to go ahead and get started?" and I'll let it get started
Once it's done, I'll ask it to run lint, typechecks, automated testing, do a code review of what's in the current git workspace, compare the changes to the spec, do my own code reviews, run it myself, whatever is needed to make sure what was written solves the problem.
In general, I'd say it's a bad idea to just let the agent go off on its own with a giant task. It should be iterative and communicative. If the task is too big, it WILL take shortcuts. You can probably get an agent to rewrite your whole codebase with a big fancy prompt and a few markdown files. But if you're not part of the process, there's a good chance it'll create a serious mess.
For what you're doing, I would likely like ask the agent to read the mega python file and explain it to me. Then I would discuss what it missed or got wrong and add additional context and explain what needs to be done. Then I would ask it if it has any suggestions for how we should break it into submodules. If the plan looks good, run with it. If not, explain what you're going for and then ask how it would go about extracting the first submodule. If the plan looks good, ask it to write tests, let it extract the submodule, let it run the tests, review the results, do your own code review, tweak the formatting, Goto 10.
I was Co-Founder and lead backend engineer for an IOS app called Welcome and our companion app Tripnotes.
- https://welco.me/
- https://www.producthunt.com/posts/welcome-2-0
- https://tripnotes.ai/
Welcome has been acquired by Dorsia in 2023, where I am a Senior IC
- https://dorsia.com
...
You can "see" my work at:
- https://welco.me/
- https://www.citizen.com/ (minor role)
- https://vimeo.com/cameo/
- https://cameo.tv/
- https://freemusicarchive.org/
- https://on-camera-audiences.com/
- https://www.laco.org/
email: [my username] at gmail
twitter: @enobrev
[ my public key: https://keybase.io/enobrev; my proof: https://keybase.io/enobrev/sigs/wEO-TfzQj7Uq10apBEHBJ150_ULPnu-Qki3JAnFy0Ss ]