最新

It's a vibe coded slop and I have not written a single line of code for this project, but I still think it's pretty impressive what the current LLM models can churn out. I used Claude Opus 4.5 for everything. The point is, I described pretty much all the features that I had in mind and it managed to implement them more or less correctly, in some instances, even surpassing my expectations. I am looking for ideas and suggestion on what could be improved and features people would like to see.

I’ve been running longer AI agent tasks (mostly in Claude Code), and I kept running into the same problem: the agent would finish or get stuck asking a question, and I wouldn’t notice until much later because I wasn’t watching the terminal.<p>So I built a small tool called Agent Reachout.<p>It lets an AI agent send me messages on Telegram when: • it finishes a task • it hits a blocker • it needs a human decision to continue<p>I can reply directly from Telegram, and the agent continues where it left off.<p>This turned long-running agent work into something asynchronous — I don’t have to babysit the CLI anymore.<p>What it does • Simple Telegram bot integration • One-way notifications or two-way conversations • Designed for “human-in-the-loop” agent workflows • Works today as a Claude Code plugin<p>Why I built it Fully autonomous agents sound nice, but in practice I often want: • approvals before destructive actions • clarification on ambiguous decisions • a quick “yes&#x2F;no” without stopping my day<p>Telegram was already where I am, so I used that.<p>What it’s not • Not a general chatbot framework • Not a workflow engine • Just a small bridge between agents and humans<p>Repo <a href="https:&#x2F;&#x2F;github.com&#x2F;vibe-with-me-tools&#x2F;agent-reachout" rel="nofollow">https:&#x2F;&#x2F;github.com&#x2F;vibe-with-me-tools&#x2F;agent-reachout</a><p>Would love feedback on: • whether others hit this problem • what notification channels would be useful next (Slack, WhatsApp, etc.) • whether this should stay a plugin or evolve into something broader<p>Thanks!

I’ve been using Cursor and Claude Code daily for real work, not just experiments.<p>One thing that surprised me is how quickly code quality converges between tools once you plan clearly. At this point, I don’t feel a meaningful difference in output quality itself.<p>What does feel different is the workflow mode each tool supports.<p>When I want many things moving at once, spawning parallel agents, delegating background tasks, or running async work, Claude Code feels more natural to me. The CLI and agent-first model fits that style well.<p>When I need to slow down, review plans, read diffs, understand context, and make careful changes, Cursor feels more friendly. It’s easier for focused thinking and sense-making.<p>So for me, it’s parallel vs focus mode.<p>We’re also starting to run Claude Code in CI&#x2F;CD for well-scoped tasks like tests, refactors, and reproducible bug fixes. That background delegation is where CLI-first tools start to matter.<p>Curious how others are splitting work between these tools, or if you see it differently.

I built PushEnv after repeatedly running into the same problems with .env files: secrets getting shared in Slack, committed to Git, going out of sync across machines, and breaking deployments with no clear audit trail or rollback.<p>PushEnv is a local-first, encrypted workflow for managing environment variables. Secrets are encrypted on the developer’s machine before being stored, and only ciphertext is ever uploaded. There’s no dashboard, no accounts, and no SaaS dependency — just a Git-style push&#x2F;pull&#x2F;diff&#x2F;history flow for .env files.<p>It also supports type-safe env validation with Zod, zero-file secret injection for CI, and versioned rollbacks.<p>This is an early version, and I’d really appreciate feedback from people who’ve dealt with secrets management at scale — especially around security assumptions, workflow design, and real-world edge cases.