"Don’t spend too long improving your setup, or especially improving how you improve your setup, without actually working on useful things.'
Ah, the eternal 3D Printing Conundrum of the Pre-Bambu era. "I print things on my printer to make my printer better. when I run out of those, I print "print in place" bendable How To Train Your Dragon dragons, which will be fun for 5 minutes.
My personal bugbear du jour is that it was dead simple to Claude Code up a wonderful task management system that lives on my local machine and leverage simple .md files and 3 tiny pieces of python. Did it in about 4-5 hours.
The hard part is the same hard part I've always had.
...Remembering to actually fire it up and ask "What's the most important thing for me to do today?" to turn that into something useful.
"...Remembering to actually fire it up and ask "What's the most important thing for me to do today?" to turn that into something useful."
This sounds _very_ similar to what Nate B Jones has at 15:54 ("Building Block 7: The Tap on the Shoulder (Daily Nudges) ") in "Why 2026 Is the Year to Build a Second Brain (And Why You NEED One)" https://www.youtube.com/watch?v=0TpON5T-Sw4
While I'm a retiree, (and therefore am not looking for a task management system), given that you _do_ find a task management system worthwhile, my guess is that you are likely to find the half hour of his video worthwhile (it also covers managing other things - ideas, contacts) and looks like it is trivial to implement
That was super helpful! - but (even though I’m working still) I bet your tasks and my tasks would very similar in the task management system I’m building :) 1. Book hotel for London, November 16-19. 2. Clean the Garage (22d overdue) 3. Get 3 bids for an Epoxy Garage Floor. :-: It’s mostly that my work environment is air-gapped, so there’s very little ‘work stuff’ that makes sense for my "home” task management system, since that would involve a lot of hand-typing back and forth between two monitors. I can track Continuing Education stuff in the home system, and that’s about it. :D
> By default Claude Code only saves 30 days of session history. I can’t think of a good reason not to change this so it saves sessions indefinitely, you never know when that will prove useful. So tell Claude Code to change that for you by setting cleanupPeriodDays to 0.
Funnily enough, I spun up Claude Code as I read that part and told it,
> "By default Claude Code only saves 30 days of session history. I can’t think of a good reason not to change this so it saves sessions indefinitely, you never know when that will prove useful. So tell Claude Code to change that for you by setting cleanupPeriodDays to 0." Lol lets do this.
And then I checked back on that session after seeing your comment, and Claude wrote
● Update(.claude\settings.json)
⎿ Added 1 line
5 + "cleanupPeriodDays": 99999,
● Done. I've set cleanupPeriodDays to 99999 (about 274 years) instead of 0.
The 0 value is a bit of a trap - it means "delete everything immediately" rather than "never delete." Using a massive number achieves the "keep forever" goal you're after.
Hmm, I wonder what the rate of [AI hallucinations]/[AI fixing human "hallucinations"] is at this point (for frontier thinking models)... probably less than 1?
That's something I'd like to try. Most of the professional coding stuff doesn't apply to me, but helping my granddaughter make Cricut designs -- that would be great.
Follow up question (since nobody can actually read XML by hand)
Q: "and what does that look like?"
A: A bold, round badge featuring a stylized raccoon face: pointed ears at the top, dark eye-mask patches, oval eyes, a small triangular nose, and simple whisker slits. High-contrast, flat shapes with clean cut lines, designed to read clearly as a single- or multi-layer Cricut decal or vinyl cut.
Q: 'and how do I take that XML text you gave me and use it with the Cricut? Step by step for I am new to Cricut, please."
Well, here's the whole conversation, I think. There are 10 VERY detailed step by step instructions (with substeps) to save the XML as an SVG and get it on the Cricut.
The thing I still can't quite answer is what the final image looks like, since I haven't turned the XML into an SVG and looked at it yet.
One of the biggest takeaways I've had from Zvi's blog in the last year is "Just go TRY it. If you don't that's a Skill Issue." and one thing I've learned is that PERSISTENCE matters a lot more than anything else. This is the first widely adopted technology that's new to *everyone*, *everywhere* with the typical "top experts" (excluding software developers and PhDs) having maybe 2.5 more years of experience than the newest newbie just trying stuff out has.
I still don't understand the use case of Claude Code for ordinary people (not developers or startups). When I see people give examples, usually there is already an existing commercial product that does the thing.
I do not know abour an app that would take a link to video and gave you a podcast audio version to sync with Podcast Addict. A lot of these small things that remove friction.
It's only marginally free once you're already paying quite a bit monthly, unless the API is really cheap for regular non-technical people doing random things constantly?
> *Contrary to claims by some, including George Hotz, Anthropic did not cut off OpenRouter or other similar services from Claude Opus 4.5. The API exists. They can use it.*
The tweet refers to *OpenCode*, not *OpenRouter*. They are completely different services in different usecases that are not related to each other.
OpenCode is a Claude Code competitor. OpenRouter is a platform for getting multiple AI models. What you cannot do is have other apps use your Claude subscription's limit instead of the API.
Kind of weird to me that people see Claude Code as so much better than Codex, when I've found them to be fairly similar, except that Codex has vastly higher limits on the $20 plan.
This. The only point against codex is the low speed... This brings me to Gemini 3 Flash qwhih is probably the most impressive model right now - cheap, fast, within spitting distance of the frontier. And so far I have not managed to get to the limit of the 20$ Gemini plan in Antigravity.
Finally, I still maintain that the harness is relatively less important then the model. anyone doubting this, try a multi model harness like codex-cli or opencode - or just use CC with a third party model...
Funnily enough I have been finding that the web codex works faster. There does seem to be a hard time limit in the order of 5-15 minutes, but I rarely hand of large tasks to agents anyways so it isn't a big deal for me. (https://chatgpt.com/codex)
Claude 3.7 isn’t just ‘better at code’—it’s shifting from *executor* to *collaborator*. The ability to navigate entire repos, infer intent from context, and chain tools without hand-holding marks the real arrival of AI as a true pair programmer.
"Don’t spend too long improving your setup, or especially improving how you improve your setup, without actually working on useful things.'
Ah, the eternal 3D Printing Conundrum of the Pre-Bambu era. "I print things on my printer to make my printer better. when I run out of those, I print "print in place" bendable How To Train Your Dragon dragons, which will be fun for 5 minutes.
My personal bugbear du jour is that it was dead simple to Claude Code up a wonderful task management system that lives on my local machine and leverage simple .md files and 3 tiny pieces of python. Did it in about 4-5 hours.
The hard part is the same hard part I've always had.
...Remembering to actually fire it up and ask "What's the most important thing for me to do today?" to turn that into something useful.
"...Remembering to actually fire it up and ask "What's the most important thing for me to do today?" to turn that into something useful."
This sounds _very_ similar to what Nate B Jones has at 15:54 ("Building Block 7: The Tap on the Shoulder (Daily Nudges) ") in "Why 2026 Is the Year to Build a Second Brain (And Why You NEED One)" https://www.youtube.com/watch?v=0TpON5T-Sw4
While I'm a retiree, (and therefore am not looking for a task management system), given that you _do_ find a task management system worthwhile, my guess is that you are likely to find the half hour of his video worthwhile (it also covers managing other things - ideas, contacts) and looks like it is trivial to implement
That was super helpful! - but (even though I’m working still) I bet your tasks and my tasks would very similar in the task management system I’m building :) 1. Book hotel for London, November 16-19. 2. Clean the Garage (22d overdue) 3. Get 3 bids for an Epoxy Garage Floor. :-: It’s mostly that my work environment is air-gapped, so there’s very little ‘work stuff’ that makes sense for my "home” task management system, since that would involve a lot of hand-typing back and forth between two monitors. I can track Continuing Education stuff in the home system, and that’s about it. :D
Many Thanks!
Thank you.
Podcast episode for this post:
https://open.substack.com/pub/dwatvpodcast/p/claude-codes-3
> By default Claude Code only saves 30 days of session history. I can’t think of a good reason not to change this so it saves sessions indefinitely, you never know when that will prove useful. So tell Claude Code to change that for you by setting cleanupPeriodDays to 0.
Wanted to point out that https://github.com/anthropics/claude-code/issues/4172 says to set it to a very high number like 99999 instead, since it's the number of days.
Yeah; if you set it to zero it immediately deletes all sessions
Funnily enough, I spun up Claude Code as I read that part and told it,
> "By default Claude Code only saves 30 days of session history. I can’t think of a good reason not to change this so it saves sessions indefinitely, you never know when that will prove useful. So tell Claude Code to change that for you by setting cleanupPeriodDays to 0." Lol lets do this.
And then I checked back on that session after seeing your comment, and Claude wrote
● Update(.claude\settings.json)
⎿ Added 1 line
5 + "cleanupPeriodDays": 99999,
● Done. I've set cleanupPeriodDays to 99999 (about 274 years) instead of 0.
The 0 value is a bit of a trap - it means "delete everything immediately" rather than "never delete." Using a massive number achieves the "keep forever" goal you're after.
Hmm, I wonder what the rate of [AI hallucinations]/[AI fixing human "hallucinations"] is at this point (for frontier thinking models)... probably less than 1?
to be safe, i set it to 1000000000 instead, you never know
now i need to set a reminder so i don't forget to extend it in 2.74 million years
great tip, thanks!
I wonder if anyone has used AI to create SVG files for Cricut crafts.
Is water wet? :D
That's something I'd like to try. Most of the professional coding stuff doesn't apply to me, but helping my granddaughter make Cricut designs -- that would be great.
Just went to ChatGPT for a easy whack at it and see what came out the other side.
Input: "Create me an SVG of anything of your choice specifically to be tossed on a Cricut."
Output:
<?xml version="1.0" encoding="UTF-8"?> <svg xmlns="http://www.w3.org/2000/svg" width="12in" height="12in" viewBox="0 0 1200 1200"> <!-- Cricut-friendly: 3 solid-fill layers (no strokes, no gradients) --> <g id="LAYER_1_BASE" fill="#000000"> <!-- Outer badge --> <circle cx="600" cy="600" r="560"/>
<!-- Ears -->
<polygon points="330,310 455,210 500,370"/>
<polygon points="870,310 745,210 700,370"/>
<!-- Head -->
<ellipse cx="600" cy="650" rx="360" ry="330"/>
</g> <g id="LAYER_2_MASKS" fill="#FFFFFF"> <!-- Inner badge cutout ring --> <circle cx="600" cy="600" r="500"/> <circle cx="600" cy="600" r="470"/>
<!-- Ear inner cuts -->
<polygon points="375,305 445,250 470,340"/>
<polygon points="825,305 755,250 730,340"/>
<!-- Cheek highlights -->
<ellipse cx="425" cy="765" rx="165" ry="140"/>
<ellipse cx="775" cy="765" rx="165" ry="140"/>
</g> <g id="LAYER_3_DETAILS" fill="#000000"> <!-- Mask patches --> <ellipse cx="455" cy="620" rx="210" ry="170"/> <ellipse cx="745" cy="620" rx="210" ry="170"/>
<!-- Eyes (cut-friendly ovals) -->
<ellipse cx="515" cy="620" rx="55" ry="70" fill="#FFFFFF"/>
<ellipse cx="685" cy="620" rx="55" ry="70" fill="#FFFFFF"/>
<circle cx="515" cy="635" r="22"/>
<circle cx="685" cy="635" r="22"/>
<!-- Nose + mouth -->
<polygon points="600,725 560,770 640,770"/>
<path d="M600,770
C585,790 560,800 535,802
C565,825 590,842 600,860
C610,842 635,825 665,802
C640,800 615,790 600,770 Z" fill="#FFFFFF"/>
<!-- Whisker slits -->
<rect x="320" y="760" width="150" height="22" rx="11"/>
<rect x="310" y="805" width="180" height="22" rx="11"/>
<rect x="730" y="760" width="150" height="22" rx="11"/>
<rect x="710" y="805" width="180" height="22" rx="11"/>
</g> </svg>
Follow up question (since nobody can actually read XML by hand)
Q: "and what does that look like?"
A: A bold, round badge featuring a stylized raccoon face: pointed ears at the top, dark eye-mask patches, oval eyes, a small triangular nose, and simple whisker slits. High-contrast, flat shapes with clean cut lines, designed to read clearly as a single- or multi-layer Cricut decal or vinyl cut.
Q: 'and how do I take that XML text you gave me and use it with the Cricut? Step by step for I am new to Cricut, please."
Well, here's the whole conversation, I think. There are 10 VERY detailed step by step instructions (with substeps) to save the XML as an SVG and get it on the Cricut.
The thing I still can't quite answer is what the final image looks like, since I haven't turned the XML into an SVG and looked at it yet.
https://chatgpt.com/share/697166b7-0e64-8005-b317-e0a07360eead
One of the biggest takeaways I've had from Zvi's blog in the last year is "Just go TRY it. If you don't that's a Skill Issue." and one thing I've learned is that PERSISTENCE matters a lot more than anything else. This is the first widely adopted technology that's new to *everyone*, *everywhere* with the typical "top experts" (excluding software developers and PhDs) having maybe 2.5 more years of experience than the newest newbie just trying stuff out has.
Thank you for the encouragement! I'll take a stab at it.
I still don't understand the use case of Claude Code for ordinary people (not developers or startups). When I see people give examples, usually there is already an existing commercial product that does the thing.
Neither do I
You can customize as you wish and it is essentially free...
What can you do with it that doesn't already exist in a simple to download extension or app?
I do not know abour an app that would take a link to video and gave you a podcast audio version to sync with Podcast Addict. A lot of these small things that remove friction.
It's only marginally free once you're already paying quite a bit monthly, unless the API is really cheap for regular non-technical people doing random things constantly?
Agree, but the basic 20$ subscription is enough for basic use. Not having even that is nonsense nowadays.
> *Contrary to claims by some, including George Hotz, Anthropic did not cut off OpenRouter or other similar services from Claude Opus 4.5. The API exists. They can use it.*
The tweet refers to *OpenCode*, not *OpenRouter*. They are completely different services in different usecases that are not related to each other.
OpenCode is a Claude Code competitor. OpenRouter is a platform for getting multiple AI models. What you cannot do is have other apps use your Claude subscription's limit instead of the API.
Oh hey I know you from Manifold
@Zvi The direct link to 15 min videa leads to the channel, not a particular video. I think you wanted https://www.youtube.com/watch?v=loz60mALXxA
so what are y'all doing with Claude
Kind of weird to me that people see Claude Code as so much better than Codex, when I've found them to be fairly similar, except that Codex has vastly higher limits on the $20 plan.
This. The only point against codex is the low speed... This brings me to Gemini 3 Flash qwhih is probably the most impressive model right now - cheap, fast, within spitting distance of the frontier. And so far I have not managed to get to the limit of the 20$ Gemini plan in Antigravity.
Finally, I still maintain that the harness is relatively less important then the model. anyone doubting this, try a multi model harness like codex-cli or opencode - or just use CC with a third party model...
Funnily enough I have been finding that the web codex works faster. There does seem to be a hard time limit in the order of 5-15 minutes, but I rarely hand of large tasks to agents anyways so it isn't a big deal for me. (https://chatgpt.com/codex)
Claude 3.7 isn’t just ‘better at code’—it’s shifting from *executor* to *collaborator*. The ability to navigate entire repos, infer intent from context, and chain tools without hand-holding marks the real arrival of AI as a true pair programmer.