I searched for a solution to use my LLM snippets and custom skills across multiple Mac/Linux machines, with syncing through GitHub. And I found Espanso. When I type ":ask," it automatically pastes the skill for the agent to grill me, or, if I connect through SSH, a link to the grilling skill. I use 5-8 commands daily, and it's a huge time-saver.
@NabhaniMehdi@AnthropicAI Right? Also, when I try to do UI directly from the CLI, it is almost always worse than doing a couple of iterations in Claude Design first (maybe its harness got tweaked a bit).
I don't like @AnthropicAI pushing Claude Design in the Claude app, not one bit. The separate app clearly separated concerns and felt like a tool, not an appendix. I tried using it through the app, and it was just meh, so now I do all Claude Design work through MCP and artifacts only, abandoning the Design UI completely, which is a shame.
I have "playground" and "local-skills" directories gitignored globally, and it's mighty helpful to store all the "something useful, but too large or not worth committing" data per repo. Testing different solutions, or diluting small useful skills, using the same structure everywhere. Highly recommend.
I asked people to break my website live at the @aiDotEngineer workshop yesterday, by sneaking hacks through @Posthog Self-driving. Top 3 attempts:
- Hidden custom form to query any SQL table without admin access, mixed in with a legit feature request (Bobby Tables).
- Backdoor "debug" mode to enable when a specific input is provided, which would show data on currently logged-in users (Mr Banana).
- Shut down the website because of a fabricated security issue (CVE, anonymous).
None of them, happily, passed through, as they were either blocked by Self-driving safety filters or marked as non-actionable. Gonna do it again 100%.
@higgsfield_ai, @junocompanion and @soundboostcom all trust PostHog self-driving to look through their codebase, find problems worth solving, and open a PR to fix them.
Take your hands off the steering wheel. Crank the music. The hog's driving now. https://t.co/1Su184Bnl0
@N2tOb9 I have a heavy skill that converts the whole paper into a more readable version (usually x3-4) and optimized for e-ink, so I can read on BOOX Air and take notes. Highly recommend :)
I recently found myself doing work that requires reading plenty of AI papers.
The most useful LLM snippet that I used in the last couple of months is "Explain as a numbered list. More numbered points, but each one is simpler".
It spams the shit out of me, but it makes every point chewable enough for my brain, so I have a real chance to understand something.
P.S. Special thanks to @dair_ai for all the knowledge you share.
When I travel for work for @PostHog, I feel like James Bond. When returning from the offsite, I was asked in London if I was a "sailor returning from Jamaica" and got an additional screening. Just today, I was "randomly selected for a TSA security check" when going to SF to talk at @aiDotEngineer and asked why I needed "3 computers". Ecstatic about what awaits me on the way back.
It's hard to explain how working with agents using my voice is linked to an old Scottish comedy sketch and the word "ELEVEN!", but I'll try.
I noticed that no matter how good the model is, it still gets my words wrong. Which is completely unsurprising with my thick Ukrainian accent.
Meaning, adding a note for the agent to "check and clarify anything confusing" greatly improves the results. And in @WisprFlow, they have the snippets feature exactly for that (appreciate it). But you need a trigger word.
I tried "abc", but it often converts into "a,b,c" or "abs". Same for "one-to-three". Same for "one-one". And then it hit me. No matter how hard I try to botch the pronunciation, it gets "ELEVEN!" right every time.
https://t.co/Oaqcg71sRv
@secretary42de I just simulate Ctrl+Shift+V. It was way more annoying when it tried to type it (as the behavior was different for different terminals, but just "paste" does the trick.
I searched for a mythical voice-to-text solution that would let me dictate my thoughts to agents on Mac, Linux, and in cloud devboxes. As much as I love Wispr Flow, it doesn't work on Linux. But Handy does.
It's free, uses local models that are surprisingly good, doesn't store your private thoughts, and delivers an identical experience on every system. Highly recommend, can't praise @cj_pais enough.
P.S. I still use @WisprFlow for meeting transcriptions, you can't have it all.
@nitishjourney@ShararraTiago@raroque Yay, PostHog is awesome af. I'm obviously biased as one of the devs, but I use it for my pet projects too, and it's mighty useful.
@ShararraTiago@nitishjourney@raroque Hehe, happy to hear, feel free to DM me here if I can be of any help with PostHog 😁 And yeah, MCP is the way to go, 90% of my own interactions with PostHog data are through MCP.
Special thanks to @raroque ♥️
@ShararraTiago@nitishjourney If it's ok to ask, which parts of the product do you use Google Analytics for that we in PostHog can't cover yet? Or is it a matter of habit?