Andrew Dunn
Back to notes
•
7 min read

Things I saved in April 2026

Here are the links and tools that caught my attention in April.

Here are the links and tools that caught my attention in April.

The control plane keeps moving up

March was the month of running ten agents in parallel. April was more: fine, but where do they live and who checks their work?

I saved Paperclip, which describes itself as “the human control plane for AI labor.” The framing is not subtle. You hire AI employees, set goals, assign jobs, give them budgets, and approve the work. One line from the page gets right to it: “OpenClaw is an employee, Paperclip is the company.”

That is the shift I kept seeing. These are less like chatbots and more like little operating systems for work. Sim is a visual workflow canvas for agents. Goose is a general-purpose local agent that can handle code, research, writing, automation, data analysis, and subagents. Twill sells the dream directly: “You push the vision. Twill writes the code, runs the tests, fixes failures, and opens PRs.”

The interface is drifting from prompt box to org chart, task queue, sandbox, and review queue.

I also saved Open Agents, Vercel’s reference app for background coding agents, and Sandcastle, a TypeScript library for orchestrating coding agents in isolated sandboxes. This is the boring part that matters. If agents are going to work on real projects, they need places to run, ways to isolate changes, and clean handoffs back to humans.

That is where the agent conversation feels more serious than it did a few months ago. Less “make me an app” and more “how do I structure a repeatable work environment where a non-human worker can do a bounded task without trashing the repo?”

Skills, specs, and other ways to make agents less weird

I saved Every’s compound-engineering plugin, which packages a workflow around brainstorming, planning, working, reviewing, and compounding what the agent learned. The line I liked was their point that the ceremony only matters if it makes the next unit of work easier.

Matt Pocock’s skills repo is even more direct: “Skills for Real Engineers. Straight from my .claude directory.” One of his skills is a grilling session, which forces alignment before the agent starts writing code. This maps to what I keep finding in practice. Claude is usually not bad because it cannot code. It is bad because it confidently starts from the wrong premise.

I saved Karpathy-inspired Claude skills for the same reason. They package habits like simplicity first, surgical changes, no assumptions, and checking the codebase before inventing a plan. This is all the unglamorous stuff that makes agent work feel less like gambling.

Ossature is another version of this idea: spec-driven code generation with tasks, dependencies, audit, and review before code. I also saved a Hacker News thread on Components of a Coding Agent, where one commenter described the magic pretty plainly: “the power that was unleashed by surrounding an LLM with a simple state machine, and giving it access to bash.” That is a good reminder. The raw ingredient is impressive, but the harness is the product.

A few saved links were about proof instead of process. pdf-proof is a Claude skill that turns AI answers about PDFs into visual proof with source screenshots. I saved it specifically for checking my tax filings.

Passmark does something similar for browser testing. Developers write end-to-end tests in plain English, the AI figures out the browser actions, and then the actions get cached so future runs happen at Playwright speed without repeated LLM calls. Stage uses AI to organize pull requests into logical chapters so reviewers can understand changes faster.

Markdown keeps eating workflows

I spent a chunk of April looking at static sites, CMSs, and content workflows because we have been working through a migration away from WordPress to a static site. That made the DemandSphere post on moving from WordPress to Jekyll with Claude Code feel like a field report from a few weeks ahead of us.

Their reasons were familiar: speed, flexibility, fewer plugin dependencies, simpler deployments, easier performance work, and less CMS overhead. Their blunt version: “The biggest issue was speed. As a platform company, we move very fast and we always felt limited by what we could do with WordPress.”

What interested me most was the implied product direction: a WordPress replacement that is basically a static site builder plus an agent-friendly workflow. If content is markdown, templates are files, and routine changes can become agent-driven pull requests, the old CMS tradeoff changes.

I saved Jekyll, Astro’s docs for building with AI tools, Astro agencies, Pooly-Press, zero-md, Screaming Frog, and Lighthouse Audit around the same time. That is a very specific cluster. Markdown, static sites, crawlers, performance checks, and agent-friendly workflows.

Kids, games, and taste

Build With Your Kid is a computational-thinking curriculum for ages 2 to 6. I love the premise. The point is not to turn toddlers into software engineers. It is to give them little mental models for sequencing, debugging, patterns, and cause and effect.

The Parent Developer’s Guide to Building Games With AI got at the same thing from a different angle. One line from the piece is great: “Then my three-year-old said ‘make me a game where a red car goes fast’ and I opened Claude instead of Unity.”

I saved Playdate, the little yellow handheld with a crank, plus the HN discussion about Playdate changing how Duke teaches game design. I also saved Microsoft MakeCode for micro:bit. These all sit in the same mental bucket: small constraints, fast feedback, playful systems.

Then at the end of the month I saved Cursor Camp, a Neal.fun game where you herd cursors around a little summer camp.

Stuff I bookmarked

A few things that did not fit cleanly but were worth saving: