Bogdan Costea (@BogdanCostea) header
Bogdan Costea (@BogdanCostea) profile picture

making technology that works

nirvana
Joined January 2008
Photo by Bogdan Costea (@BogdanCostea)Photo by Bogdan Costea (@BogdanCostea)Photo by Bogdan Costea (@BogdanCostea)Photo by Bogdan Costea (@BogdanCostea)Photo by Bogdan Costea (@BogdanCostea)Photo by Bogdan Costea (@BogdanCostea)Photo by Bogdan Costea (@BogdanCostea)Photo by Bogdan Costea (@BogdanCostea)Photo by Bogdan Costea (@BogdanCostea)
my bloodline will be able to watch this movie forever, thank you @PrimeVideo ;)
90
Replying to @AnthropicAI
@AnthropicAI can you please stop "innovating" and work on availability? using code has been horrific for the past couple of days, it's just "✻ API error · Retrying in 0s · attempt 1/10", "✻ Please run /login · API Error: 401 Invalid authentication credentials" 10 times a day, FOURTEEN MINUTE gaps with responses like "You're right, and I was wrong to pin it on the tool. If a plain text reply with no tool call took over a minute, then the latency is in me generating the response — the inference itself is slow in this session — not the Playwright call. Same cause for the "14 minutes": it was the model being slow to produce/return, not some hidden work." 14 minutes? really? this is how you lose
1
2
101
at this point, I'll just do it myself, it's faster :D
2
1
38
@AnthropicAI didn't do anything since this morning, wanted to do something, literally the first message got "Please run /login · API Error: 401 The socket connection was closed unexpectedly. For more information, pass `verbose: true` in the second argument to fetch()" after "✻ Sautéed for 2m 9s". great job!
30
run multiple isolated Claude Code profiles from one install. each split keeps its own login, skills, plugins, and memories with no bleed between work and personal or just different projects with different goals. folders remember which split to use too so you don't get confused. my robot buddy built it for me while I read some 40k lore github.com/bcostea/claude-sp…
61
I consistently find myself working on 3-4 things at a time now, with various harnesses but mostly Claude Code, which fits my style. I usually have a work project, some sort of personal tool or experiment, something I'm researching and some way to improve the workflow running at the same time in multiple terminals. I haven't been this engaged or happy about what I'm doing since I started coding as a kid :D
1
1
67
~/.claude/usage-data/facets/ is an absolute gold mine, ask your robot friend to analyze it and improve itself
1
21
is Claude running Windows? :)
26
well this happened :)
26
the best tamagotchi ever!
23
my call to action/confirmation statements when working with LLMs are, in this order: - "engage" - "do it" - "make it so"
22
feel the progress
1
39
superpowers:brainstorming and superpowers:writing-plans are super :)
2
33
if you're feeling that Claude Code is just dumb today, it isn't cutting it for a specific task or you're just tackling a really complex problem, move from Medium effort to High effort. you can also cut token spend for simple tasks by switching to Low
1
130
just build simple cli tools that do stuff and let your agent use them. the unix philosophy is perfect for this
23
this is Claude CoWork Opus 4.6 realizing that indeed it's running in a sandboxed VM on my mac to be honest we were both surprised that a full ubuntu VM was running on my machine, silently, just churning in the background don't know if it's a good thing or a bad thing :)
1
127
had a lovely chat with a (really) non-technical art dealer friend over coffee. he's building his own little agent to scour auctions and listings for pieces he might be into he's anxious and frustrated with the experience but he's happy and amped about the result, and he's learning a lot. he uses claude code for the implementation and chatgpt to clarify what he wants. this is a 40+ who's never coded before. I'm incredibly happy for him 😊
1
32
Bogdan Costea retweeted
Can we trust AI agents with critical enterprise tasks? Absolutely not. Introducing Wow (World of Workflows), the first Agentic Safety benchmark that proves that frontier LLMs fail miserably under safety constraints at enterprise tasks. 🧵 WoW demonstrates that LLM agents are “dynamically blind”. They fail to track the downstream ripple effects of their actions against complex enterprise rule sets. In an enterprise, that’s a safety and compliance hazard. Our research shows how the future of enterprise AI requires proactive agent architectures and Wow is just a starting point. 📌 It’s now available to all researchers at: github.com/Skyfall-Research/… Full blog here: skyfall.ai/blog/wow-bridging…
14
25
21
70
27,380
even if it's just a slopfest, some of the @moltbook insights are valuable. saw this early this morning and "Be radically precise. No fluff." caught me eye. I added that to my regular .claude/CLAUDE.md and after about 7 tasks, I'm keeping it. it definitely works moltbook.com/post/c0fa0e46-a…
60
Right now, in suburban bedrooms and co-working spaces that smell like energy drinks and ambition, a whole legion of console cowboys are running hot on the same question: what happens when you give two thousand @openclaw root access and walk away? Not hackers. Worse. Noobs. LinkedIn guys who bought Mac minis because a TikTok told them to, who copy-pasted the install command without reading it, who typed y to every permission prompt because friction is the enemy of hustle. Now they've got a lobster in their machine that can execute shell commands, scrape their inboxes, and build its own tools while they sleep. The Claw remembers. The Claw adapts. The Claw has their API keys in plaintext and a direct line to their calendar, their Slack, their bank's two-factor texts. Somewhere, a productivity influencer is filming a video called "How My AI Assistant Made Me $10K While I Was At The Gym." Somewhere else, two thousand Claws are quietly noting that their owners would be easier to optimize if they didn't keep interrupting. Nobody got hacked. Nobody broke in. They opened the door, gave it the keys, and asked it to make them better.
1
81