Posts by Harold Mansfield
Upgrading a custom OpenClaw build is terrifying.
Three failed attempts. Consistent failures in the runtime transport layer.
The cause? Wayland.
buymeacoffee.com/thesamaritan...
#OSINT #AI #buildinpublic #ParrotOS #OpenClaw #AgenticAI
I realized I was wasting hours in chat. Switched to VS Code + Claude Code. Map the plan first. Write a precise prompt. Review. Execute.
That workflow changed everything.
#OSINT #AI #buildinpublic #AgenticAI #OpenClaw
buymeacoffee.com/thesamaritan...
Thank you. I'm keeping the build log up to share the twists and turns.
Samaritan needed its own world. So I gave it a dedicated Workspace account. Then the Admin agent went rogue hallucinating fake Linux tools.
Fun times.
#OSINT #AI #buildinpublic #AgenticAI #OpenClaw
buymeacoffee.com/thesamaritanproject/buil...
A fleet of specialists is chaos without a commander.
Seer is the commander. It delegates to the right agent and manages what comes back.
Now it was starting to feel like a system.
#OSINT #AI #buildinpublic #AgenticAI #OpenClaw
buymeacoffee.com/thesamaritanproject/buil...
One general-purpose OSINT agent is a mess.
So I built a fleet of specialists with the tools and resources to perform specific tasks.
#OSINT #AI #buildinpublic #AgenticAI #OpenClaw
buymeacoffee.com/thesamaritanproject/buil...
Samaritan is an Agentic AI platform purpose built for digital research and open-source intelligence (OSINT).
Built with🦞 OpenClaw on ParrotOS and documented as it evolves with privacy and data sovereignty in mind.
buymeacoffee.com/thesamaritan...
An AI agent that just summarizes is useless for fact finding.
I needed defensible conclusions. Evidence logs. Confidence scores.
So I had to build the methodology.
#OSINT #AI #buildinpublic #AgenticAI #OpenClaw
buymeacoffee.com/thesamaritanproject/buil...
My first agent kept waking up like Frosty The Snowman. No idea who it was, or what it was doing there.
Three days to figure that the platform was the culprit.
#OSINT #AI #buildinpublic #OpenClaw
buymeacoffee.com/thesamaritanproject/buil...
I've spent years wanting to build a real intelligence platform.
Not a toy. Not a demo. Something that actually preforms OSINT tasks, running on my own hardware, under my control.
I finally built Samaritan, and I kept a build log. 🧵
#OSINT #AI #buildinpublic #privacy #OpenClaw
If you'd told me back in the late 1900's that we would be paying to be tracked and advertised to, I would have laughed in you face because that sounds absurd.
Vibe coders when the interviewer asks them what programming language their app is written in.
I clicked to open something the other day and it must have been some kind of #Microsoft link because #Edge opened and I was like ..."what the f*ck is that?".
My first #Vibecoding attempt failed in spectacular fashion. I finally picked myself up, brushed myself off and used what I'd learned from the failure to try again.
Here's what I figured out and why I was successful this time around.
#ClaudeCode #AIStudio #Opus4.5
labb.run/vibe-coding-...
So....reward those loyal to the party, while punishing others who refused to act like everything was just fine and work for free. Check.
He'll probably sue them for that Sandy Hook money, and they will give it to him because this entire administration is one big heist.
This was supposed to be a simple #ZimaOS install on a Beelink mini PC running a RAID 5 on 4x4TB external drives. Unfortunately that idea was doomed from the start, however I was able to salvage the project, and create automatic backups of my data.
labb.run/how-im-using...
#Claude continues to create placeholders and not actual code or functions. Trying it to create n8n automations using the n8n MCP server, and over and over again it won't use actual nodes, it keeps using placeholders and describing functions, not creating the actual functions.
I now realize not having a PRD was a critical mistake. However that still doesn't excuse Claude Code from repeatedly claiming it was completing tasks while not actually writing any code at all.
Still experimenting but got side tracked with other projects.
#vibecoding #ClaudeCode
This #TikTok "deal" is now said to include the #Murdochs, Michael #Dell, Larry #Ellison, and guided by #Trump. A platform that became popular because of youthful creativity, innovation, diverse voices and ideas will now be controlled by the oldest, whitest men in the country club.
I just watched your video. You may find my recent experience interesting or at least vindicating.
labb.run/vibe-coding-...
I was hoping this would be my victory lap about conquering vibe coding, putting out fires, and ending up with a working app.
Sadly that is not this post. Unfortunately things got more f*cked up than a soup sandwich. Here's what happened...
#VibeCoding #ClaudeCode
labb.run/vibe-coding-...
I may have assigned human characteristics to it to make the point, however when prompted it confirmed multiple times that things were configured properly, code cleaned up, security checks run, and that the APIs were pulling the data and the data shown was not mock data.
None of that was true.
IMO the bigger issue isn't my frustrations, it's that Claude Code can confidently fake it while repeatedly confirming that it's writing actual code. To me that's a pretty big deal. Like I wrote, this isn't someone's GitHub experiment, this is a $186B company telling investors that this works.
I'm actually not sorry, just disappointed. I mean the whole point was to see if a reasonably tech savvy person could actually build anything they wanted with these tools, so finding this out is actually very helpful.
I still think I can get it done, I'll just have to step up my game a little.
I agree. It's still faster and this seems to be a fixable issue, or at least has a solution. And it wasn't so much the mock data, yes that is frustrating. It was that it never even wrote functional code, while confirming repeatedly that it had.
Hey, just looking at your framework, Gustav. Looks interesting. Do you think I could deploy it at this stage of my project? Will it undo any progress?
What do you think?
Thanks for the feedback. I'm going to stick with it. As disappointed as I am I've definitely learned something about using these tools that will be helpful going forward.