@nikivdev
nikiv.dev
Joined May 2013
  • Tweets9.5K
  • Following7.5K
  • Followers5.1K
1.8KMedia
my new dev flow for more complex things is i have codex write spec claude review it codex review what claude wrote, optimize spec get review fast (will optimize this with tooling, wip) then i have cli `spec exec <.md path>` and it writes optimized prompt and executes it
3
14
3,444
soon `spec review <.md path>` will open review tool with inline comments basically github light will be under unhash.sh/<hash> all inline all reactive ai will be able to fetch optimized review feedback with context fast to iterate on spec
1
4
1,877
Truly life changing day @tekezo is genuinely top 3 humans I admire the most who changed my life Since this post, I wrote PR to Karabiner with the change to cut latency and remove shell to use sockets I thought I would have to support a fork but no, tekezo literally merged it
nikiv.dev
@nikivdevFeb 8
Started replacing keyboard maestro with my own macro engine in c++ and it's crazy how doing this moved my most common op to open apps from 500ms in KM to 200ms in my own engine I am now evaluating forking karabiner elements to avoid even more overhead on engine calls
2
13
2,319
more replies
the only way to get these lower is joining macOS team and improving things there + better hardware but this is plenty fast and at current physical limit which feels amazing on side I made seq support notion of `.md` executable recipes which is super cool
2
1
763
tested it as i had around 30+ zed windows open and I wanted full restoration I had my mac literally go tab by tab, full snapshot of all tabs (all claude/codex sessions inside each window) the stuff that seq lets you do is absolutely insane tbh
1
667
Started replacing keyboard maestro with my own macro engine in c++ and it's crazy how doing this moved my most common op to open apps from 500ms in KM to 200ms in my own engine I am now evaluating forking karabiner elements to avoid even more overhead on engine calls
2
1
9
3,379
more replies
not crazy enough to swap karabiner on my main mac so seq now has a concept of an `action pack` which is series of seq instructions that can execute if i pass the pack signed over tailscale network daemon will instantly execute it if its set for execution will harden it more
1
552
idea behind action packs is that i can remotely run kind of like executables fast from my main mac on another as there is efficient vm running that can process it all if it works well for this case, will exp with it by running seq in linux hosts and agentic containers too
1
399
Here's the best blog post I've read this week, by the @dottxt "The next generation of LLM systems won't be judged by how well they format output. They'll be judged by whether they can be trusted to act."
.txt@dottxtaiFeb 4
We spent two years getting LLMs to speak valid JSON. That was the easy part. Now we're solving what actually matters: making sure the model can't even express an action it's not allowed to take. Not "will not." Cannot. 👇 blog.dottxt.ai/control-layer…
1
4
28
3,010
my `f commit` system that uses agent to review a diff before it commits and pushes it now also ends commit with a hash that encodes ai chats it took to create the commit + lessons/skills that would help to be encoded to not repeat same mistakes + todos with improvements tracked
1
1
3
729
more replies
trying to cut vscode fork to the point where iterating on it becomes quick I can already always do `f sync` to always keep up to date with upstream which is lovely even as I make heavy changes to the code I will have agents with rules to know what to do on `sync`
2
346
but so far I just cut all linux/windows targets etc as I don't care about those and trying to make vscode into a proper diff/commit review tool with better AI I have the infra in place for this with hashing, just need to wire things up well
300
Load more