Here's what I've been watching: startups do the market research, build the product, create the category and Claude Code ships the exact same feature once the market adapts

So here's what I think the next 6–12 months look like:

1. Self-Evolving Agents: Skills exist today as a framework. Agents will begin updating their own skills, rewriting their own playbooks after every task. The system learns to improve itself. Lots of startups already working to solve this problem

2. Local → Remote → Global Sessions: They shipped local where an engineer works on few repositories within Claude Code. Then recently remote to push local sessions to remote. The trajectory seems obvious. Soon, the entire organisational codebase becomes a few global sessions tended by few architects within the organisation

3. Moving into the Production Layer Claude Code moves from the editor into the infrastructure itself watching, identifying bugs, shipping fixes.

4. [2026] The Great Reallocation This is the hardest one to say. Companies won't just hire fewer engineers. They'll restructure around the assumption that the system runs itself with fewer engineers.

I would love to be proved wrong!!

A developer's machine accumulates tools fast. A Rust CLI you compiled last year, a Python formatter installed via `uv`, a language server pulled from npm, a terminal emulator from a curl script, a Go binary built from source. Each came from a different package manager, each with its own install incantation you half-remember.

I wanted a way to declare what I need without adopting a complex system like Nix or Ansible just for a single laptop. The result was a plain old Makefile.

I wrote a short post on using Make (along with a tiny bash script and fzf) to create a searchable, single-command registry for all your local dev tools. It’s not a new framework or a heavy tool—just a simple way to organize the package managers we already use.

If you're tired of losing track of your local environment, you might find it useful.

2 computers updated to Tahoe in the last days, both did not accept the user password after that and both hat to do resetpassword from recovery mode.

Happened to anyone else?

I was frustrated seeing high-end servers struggle with tasks that a $100 phone could handle with the right architecture. So I built this.

Just as the title says, isn't it stupid to run AI in a production environment before we have addressed some pretty major fucking problems with it?

Anthropic says on their safety page: "these filters are not a security boundary." Snyk found 36% of ClawHub skills contain security flaws. Trend Micro documented malware being distributed through ClawHub.

These agents have shell access, file access, and connected accounts. We built Shoofly to sit in front of tool calls before they fire.

- PreToolUse / PostToolUse hooks intercept every tool call - Blocks prompt injection, credential theft, unauthorized writes, malware in tool results - Works with OpenClaw, Claude Code CLI, and Cowork / Dispatch - Open YAML policy -- read it, fork it, audit it - Free tier detects. $5/mo blocks.

The Cowork piece was the interesting part. Cowork runs Claude Code inside a full Ubuntu VM -- host hooks don't fire there. We used the plugin system with hooks/hooks.json and VirtioFS to get sub-50ms alert latency from inside the VM to host notifications.

curl -fsSL https://shoofly.dev/install.sh | bash

shoofly.dev

Slow setup? Heavy deps? Complex configs?

What if visual testing took 5 minutes?

What's blocking you from testing UI changes?

QR codes are convenient, but they have a fundamental weakness: they can be easily reused by copying or taking screenshots. To solve this, I built a system that does not change the QR code itself, but dynamically controls the meaning and validity of the scan result on the server side. With the same QR code, it is possible to enforce: • One-time use • Time-based expiration In other words, the QR code remains the same, but reuse can be prevented. For example, instead of generating multiple QR codes, a single QR code can be used to control multiple products or items. Here is a live demo (30 seconds): https://dynokey.ngrok.app/ The first scan works, and the second scan is blocked. This approach can be applied to ticketing, access control, authentication, and more. Patent pending. If this is useful, feel free to reach out: info@xinse.jp Feedback is also welcome.

ProofShot = 100MB + complex setup.

I just want screenshots + diffs + reports.

Built a 100-line Python version in 30min.

Who else hates heavy testing tools?