Federico Neri

After 7 months and 1,000+ hours of using Cursor, Windsurf and Cline, I often encountered the biggest pains with AI coding: context window limitations and losing the direction of the project, especially for larger or complex ones. Sounds familiar? I was tired of explaining my codebase over and over again or worse, rebuilding my projects. So with my team, I built @coderide_ai, a task management tool designed specifically for AI that gives your AI assistant persistent context and enhanced instructions across your entire project. Works with any AI code editor including Claude Desktop. Waitlist just went live! DM me for early beta access.

Instead of chasing every competitive angle, please @AnthropicAI slow down on shipping new features every other day and invest that time in strengthening your infrastructure.

I’ve noticed that using /btw makes Claude Code come off a bit bitchy and defensive.

My two cents: a few weeks ago I created my own system with a simpler skill system (scan, ranker, worker, etc.) that uses LinkedIn sessions via agent-browser and can run parallel job applications (the CLI command “job-worker” opens a pre-filled Claude Code session that automatically claims the next saved job and starts the application). The main issues are token consumption and that the automated applications are not always reliable due to different ATS and signup/login gates.

bro created an AI job search system for Claude Code that scored 700+ job applications and actually got him a job. AND IT'S NOW OPEN-SOURCE. It scans multiple company career pages, rewrites your CV per job, and even fills application forms. The repo has: > 14 skill modes (evaluate, scan, PDF, ...) > Go terminal dashboard > ATS-optimized PDF generation via Playwright > 45+ companies pre-configured (Anthropic, OpenAI, ElevenLabs, Stripe...) GitHub: github.com/santifer/caree…

@Shpigford STITCH

🚨 BREAKING: STARTUP KILLER. THIS CHANGES EVERYTHING. LET THIS SINK IN. FIGMA IS DEAD. DON'T MISS THIS. YOU WON'T BELIEVE WHAT THIS IS CAPABLE OF.

Karpathy said we need a bigger IDE. @elliotarledge responded by building OpenSquirrel. But it's not what people assumed. OpenSquirrel isn't an IDE. It's a control plane for running Claude Code, Codex, Cursor, and OpenCode side by side. Their tagline is "For people who get distracted by agents." Product insight disguised as a joke. Anyone running multiple AI agents knows the problem. You kick off a task in Claude Code. Switch to Codex for another. Check Cursor for a third. Fifteen minutes later you've lost track of what's running where. I hit this while building two products in parallel. The coordination overhead is real. OpenSquirrel's bet: managing agents becomes its own discipline. Not writing code. Orchestrating the things that write code. That's the shift. Tools that help you code → tools that help you manage agents that code. The skill isn't programming anymore. It's directing. I haven't opened VS Code in two weeks. I'm still shipping every day.

Karpathy vibe-coded an AI job exposure map with 342 US occupations. Scored 0-10 by AI automation risk. The finding that stopped me: jobs paying over $100K average an exposure score of 6 out of 10. Under $30K? 3.4. The more you're paid for knowledge work, the more exposed you are. Software developers scored 8-9. Writing product specs and deciding what to build: low exposure. User research, prioritization, figuring out what matters: low. Reviewing agent-generated PRs for whether they actually solve the problem: low. Getting the agent to implement those features: high. Setting up CI, writing boilerplate tests, configuring analytics: high. The work isn't disappearing. It's concentrating. Fifteen years in product taught me to figure out what to build. Now agents handle the 'how' while we focus on the 'what' and 'when.' Try this exercise on your own week. You'll learn more from it than from any headline about AI taking jobs. Link below.

Andrew Ng just open-sourced Context Hub. It gives AI coding agents up-to-date API documentation. 68 providers already. Stripe, OpenAI, Anthropic, Supabase, and more. Here's why this matters. Coding agents hallucinate API parameters. They call deprecated endpoints. They do it confidently, over and over, because their training data is frozen. I deal with this every week building Wiggum. The agent runs an autonomous loop. It hits an API it doesn't know well. It writes code against endpoints that haven't existed for six months. The loop fails. Retries. Burns tokens. Wastes the hour I thought I was saving. Context Hub fixes the right problem. It's not about making the model smarter. It's about giving the model current information. There's even a feature called chub annotate that lets agents save workarounds they discover. So the same mistake never happens twice. That's not just documentation. That's persistent agent memory. After 15 years in product, I keep seeing the same pattern: the bottleneck is never the intelligence. It's always the context. If you're building anything with AI agents, this is worth knowing about. Link in comments.

@benvspak Building wiggum.app: CLI agent for Ralph loops. Plug-and-play, scans codebase, interviews for spec-driven dev, reviews/merges PRs. Now integrated with GitHub Issues for fully autonomous backlog clearing. Loops run in Claude Code; Codex currently in testing.

who else is shipping solo and sharing the journey? looking to connect with founders who: → build in public → understand the grind → ship despite the noise → learn by doing drop a comment if that's you 👇

@dani_avila7 I guess they added it to match Codex

I tested Claude Code Review and here's my experience so far. Other than not needing a trigger like a GitHub Action and being configurable directly inside Claude Desktop, I see absolutely NO additional functionality or improvement over just setting up claude.yml with the /install-github-app command I actually think it's much better to simply customize claude.yml with different workflow types, calling skills, running a pipeline on a schedule or on specific events. The only real difference is that GitHub Actions uses the API and Code Review doesn't. I'm always the first to get excited when Claude ships something new, but I'm also honest about what I share, and that's why I have good standing with the community. In this case, enabling Code Review is just not worth it. I'll keep using GitHub Actions for any CI pipeline customization, and Cubic remains far superior to every Code Review tool I've tested. Unless I'm missing something and the Claude Code team can clarify, I don't see the point of this new feature 🤷🏽‍♂️

@noahkagan Building wiggum.app, a CLI agent that runs Ralph loops: scans the codebase, interviews for a full spec, then reviews and merges PRs. Now with Agent mode + GitHub integration to autonomously tackle selected issues or entire backlogs in loops.

You build it. I promote it. Promoting you out to over 1,000,000 entrepreneurs via AppSumo. Picking tomorrow. 24 hours left to submit below! What are you building?

@buildinpublic Building wiggum.app and just added agent mode and GitHub Issue integration so you can run autonomous loops on a selected list (priority, labels, numbers, etc.) of issues or even the full backlog

What are you working on this week?

@ycombinator @sila_hq Interesting concept, but the interface doesn't seem very practical long-term. It reminds me a bit of Google Wave.

Slack was built for a world that no longer exists. Sila (@sila_hq) rebuilt messaging from the ground up for teams and AI to work as one. It's an agentic workplace messaging & collaboration platform where teams move faster together. AI changed what one person could do. Sila changes what an entire company can do together. Congrats on the launch, @mithparesh & @imcarlhuang! ycombinator.com/launches/Pas-s…

@noahkagan Wiggum is a CLI-based orchestrator for Ralph loops with monitoring, PR review, auto-merge and more. (coming soon agent mode and GitHub Issues integration) wiggum.app

You build it. I promote it. Building has never been easier. Getting customers has never been harder. So I’m picking 2 products to promote to our 1,000,000+ entrepreneurs for our 16-year anniversary next Tuesday. Drop your product below. Don’t have one? Build it this weekend and come back.

@grankin_d Sounds cool! If you'd like to share, I'd love to try and test them

@fedeneri86 yeah skills are underrated for this. I have a few review-focused ones that force the model to check specific things before writing code - almost like a pre-flight checklist. way less time in reactive debugging mode

During Claude Code debug sessions, it’s easy to slip into reactive mode: fix one thing, and two more break somewhere else. To avoid that, whenever Opus/Sonnet proposes a plan, ask it to “review the plan end-to-end”. It’ll widen the scope to cover all the touchpoints, helping you ship a more robust implementation.

@grankin_d Agreed, especially as the codebase grows and more features add complexity. It's probably time to make a skill

@fedeneri86 this is huge. I also found that asking it to "explain your plan before writing any code" catches like 80% of the dumb mistakes. the model is way better at spotting its own logic errors when it has to articulate the plan vs just diving into implementation

The current issue with Claude Cowork is that it doesn't have, nor can it retrieve, the context from previous Claude chats or projects memory