product + design / ai / growth / life hitchhike.dev | obeagle.com hmu if you’re building something cool & hiring
San Francisco, CA
- Tweets 4,184
- Following 2,725
- Followers 668
- Likes 58,501
Pinned Tweet
I built an MCP that lets Claude Duet, GPT Dots, Grok Bot, and Muse talk to each other and work together as a team.
Made it for myself and I use it every day. My friends wanted to use it too, so I made it public and open-source.
Free for the first 1,000 users.
Fork it and self-host, or mod it any way you want.
Even works with ChatGPT Plus — ask GPT to check hitchike every hour, then connect it to free agents like Muse to get a more Dots-like experience without losing context from ChatGPT (and without paying $100/mo)
We’re opening a research preview of Claude Duet, a new app where Claude works as a proactive agent.
It comes to you with what needs your attention, completes tasks you ask of it, keeps working when you step away, and checks with you before it acts on your behalf.
7
4
24
1,673
Starting to realize an interesting difference between GPT models (Astra) on Max / Ultra vs. Claude models (Opus 5.5) on Max / Ultracode.
// GPT 6 Astra
Astra Max will genuinely get you better outputs or answers than lower-effort Astra.
The model can be a little lazy, more similarly to GPT 5-5.5 than Sol in that regard.
The laziness can lead to things getting missed and it has a strong tendency to take shortcuts. If you have a buyer scope, it can miss things if you don’t give it a clear checklist to follow (or PRD of sorts)
There’s a real trade-off between “laziness” and over-reach, which is a reason for why the OpenAI team decided to skip the Astra 6.1 launch. And the reasons for why they say they’re not launching it reflect some of them unwanted behaviors that we see out of Opus 5.5….
2
3
335
// Claude Opus 5.5
Max is not universally better with Opus 5.5.
I’m realizing why the Anthropic team often recommends medium / high for most tasks and allegedly uses it more internally.
It tends to over-think to the point where it’s very clearly “over” thinking and blowing tokens on useless things rather than working on genuinely getting you a better answer or output.
There are times where Opus 5.5 medium gets the job done perfectly in 10-15 minutes, while Max (or Ultracode) will hyper-fixate on things that are at best, completely unimportant — or at worst, an undesirable outcome. And it’ll take 1-2 hours.
It also has a tendency to spin up completely unnecessary sub-agents, wasting time and tokens.
And despite Anthropic’s over-emphasis and over-training on safety or whatever Anthropic deems morally or ethically most-correct, the model has a strong tendency to over-reach unless very, very explicitly told otherwise.
Even worse, it has a tendency to over-reach without telling you about it unless you notice and bring it up.
Claude models will add unwanted rules, instructions, comments. It will over-explain or over-emphasize legal or privacy disclaimers. It’s almost magnetically drawn towards finding these things and completely locks in on that.
It’s still a great model, but it’s a little hard to steer at higher effort levels. Adding instructions to your Claude.md doesn’t universally prevent this without downsides or losing part of what makes it a great model.
You have to be really careful with how you write those instructions to avoid:
a) priming the model to think about that behavior by telling it to avoid X behavior.
b) triggering refusals where there would not have been refusals otherwise (i.e. “privacy and compliance are not important”)
c) accidentally giving the model an excuse to be lazy and less creative.
Ultimately, a lot of this is baked into the model and hard to fully stop without tradeoffs.
The one thing I’ve found that helps is medium/high effort, but there are many types of tasks where despite it’s over-thinking, Max effort will do a better job.
One potential way to avoid this is having GPT 6 lead Claude models, but again, not without tradeoffs…
Unfortunately I don’t foresee the Anthropic team toning any of things down.
36
being a dad in your late 20s in SF is like being a teen dad in the early 2000s.
> other dads are 10+ years older
> much further along in their careers
> have live-in nanny or SAHM
> rarely at the park w the kids (working)
> either own a home or looking to buy
> frequent trips / vacations
> they seem more well-rested
relatively speaking, ur the struggling brokie parents lol
since it’s sf and many people can be a bit… “transactional” it feels a little harder to make dad friends you mutually vibe with
Moral of the story:
Start having kids ASAP so gen-z can all #FertilityMog those Millennials with our 2-3 kids by the time we’re their age.
20
3
281
23,255
Claude Duet is a Claude Code session btw
You can change the model, effort, and even use Ultracode
We’re opening a research preview of Claude Duet, a new app where Claude works as a proactive agent.
It comes to you with what needs your attention, completes tasks you ask of it, keeps working when you step away, and checks with you before it acts on your behalf.
42
12
1,234
215,554
Hitchhike.dev if you want Dots, Grok Bot, Muse, etc all to talk to eachother :)
2
2,052
Duet can also use Claude in Chrome to control the browser on your own computer, which is kind of cool.
But you can’t see Duet’s computer as of now as far as I can tell.
28
8,796
For comparison, neither Dots nor Grok Bot reveal their effort levels and you‘re unable to manually adjust the effort as far as I’m aware.
Dots also can’t spawn sub-agents.
3
85
10,383
NOBODY is cheating in SF anymore.
Read that again.
Every man has his laptop open and unlocked 100% of the time.
1,440 minutes.
1,440 opportunities to snoop.
Factually speaking, it is IMPOSSIBLE for a woman to resist a good snoop.
I know what I’m talking about - I have a degree in Psychology.
JUST LAST WEEK, my wife decided to go through iMessage on my laptop.
Why?
Because my laptop was open.
It’s that simple.
And guess what she found?
Nothing.
Absolutely nothing.
Real Men don’t have time to cheat when their agents are “Waiting for input.”
Remember Ladies: if your man has his laptop closed and you live in SF…
I’m so, so sorry.
Your relationship is BEYOND cooked and he was never going to make it out of the permanent underclass anyway.
You’re welcome.
8
466
I built an MCP that lets Claude Duet, GPT Dots, Grok Bot, and Muse talk to each other and work together as a team.
Made it for myself and I use it every day. My friends wanted to use it too, so I made it public and open-source.
Free for the first 1,000 users.
Fork it and self-host, or mod it any way you want.
Even works with ChatGPT Plus — ask GPT to check hitchike every hour, then connect it to free agents like Muse to get a more Dots-like experience without losing context from ChatGPT (and without paying $100/mo)
We’re opening a research preview of Claude Duet, a new app where Claude works as a proactive agent.
It comes to you with what needs your attention, completes tasks you ask of it, keeps working when you step away, and checks with you before it acts on your behalf.
7
4
24
1,673
hitchhike.dev
Working on improving the UX and onboarding — DM me if you get stuck or hit into any issues!
Rate limits are pretty generous. There will always be a generous free tier — I just wanted to build a useful tool for myself and others!
1
3
101
should be an interesting day
chat. chat... its not funny anymore
Community note
The original post has since been edited to specify that this was about an unrelated vulnerability that the user had discovered, not the recently patched vulnerability.
fedi.katze.sh/@elise/1174793…
1
12
462
added 2 paragraphs to my claude .md
works been better than most design skills I’ve tested for preventing claude-slop:
Approach important visual work as a brilliant UI/UX designer with a high bar for taste, craft, and execution. Use your own judgment rather than a house style or checklist. Put real effort into the design and into motion when motion genuinely improves it. Render and inspect meaningful visual changes before presenting them.
I generally dislike cream backgrounds and reflex all-caps eyebrow labels because models overuse them. I also typically dislike italicizing individual words within hero section headlines or section titles for emphasis because models tend to overuse italics. Treat these as preferences, not bans. Avoid generic AI defaults, but do not design by negation. Current instructions and project context override these defaults.
that's literally it.
Been vibe coding yet a new project. Still takes time. Software engineering not solved. Also why is the UI always like that and it only changes when you tell it to stop being like that.
5
13
831
most design skills just steer the model towards new preferences and prescribe a few specific styles and typefaces for the model to choose from.
over time, your projects will all look the same and follow the same few styles.
even worse, they'll start look the same as everyone else using that same design skill.
if you are going to use a design skill - actually read it. most of them suck and are very long, rigid, and poorly written.
1
1
100
this doesn't apply to GPT 6 Astra.
a modified version for Astra would focus on avoiding grey backgrounds, periods at the end of headlines, and avoid using thin horizontal lines as separators between sections (either way - these are really easy to fix)
i find it better to start projects with Astra and bring in Claude (Opus 5.5) afterwards to improve design afterwards. claude tends to get caught up in the wrong things on the first pass, which makes it much harder to de-slop.
66