The Five I Actually Run

I do not recommend tools because they trend. Most "best plugins of 2026" lists are stale by the time you read them, and they all name the same things. This is a different list. These are the five I open every day, and each one is a principle this blog has argued for, now shipped as software you can install.

That is the whole filter. Not popular. Installable conviction.

Superpowers, by Jesse Vincent. An agentic skills framework that enforces a real workflow: spec, plan, test-first, systematic debugging. It ships as a Claude Code plugin and works across other agents too. This is the spec-driven loop as installable software. I run it because it holds the discipline on the days I would skip it, and those are the days that need it.

claude-mem, by Alex Newman. A persistent-memory plugin. It captures what the agent did in a session, compresses it, stores it locally, and feeds the relevant parts back into the next session. Context survives across sessions and across compaction. This is fix the system, not the output made automatic. The decision gets remembered without me re-explaining it tomorrow.

Obsidian. A local-first Markdown notes app where a vault is a plain folder of files an agent can read directly, no export. This is your AI context belongs to the team, the durable knowledge base under everything. Honest note: there is no official MCP server, the integrations are community-built. The plain-folder format is the reason it works anyway. The model reads the same files you do.

Playwright MCP, by Microsoft. The official MCP server that lets an agent drive a real browser through structured accessibility snapshots instead of screenshots. This is the record-then-refine half of trust the pipeline. The agent works the actual UI, and you keep the trusted end-to-end tests that come out of it.

ui-ux-pro-max. A UI/UX design skill: styles, color palettes, font pairings, layout and interaction guidance the model draws on when it builds interfaces. This is the taste a raw model lacks, the gap don't just use the model named. It is opinionated, and that is the point. Defaults beat a blank page that fills with generic AI slop.

What this is not

• Not a buyer's guide. These are the ones I run, not a ranking of the field. I left out good tools because I do not use them daily, and a recommendation I cannot stand behind is worth nothing.
• Not permanent. This list is true today. Tools in this space turn over fast, and the principle outlasts any of them. If a better one ships next month, I will swap it and the principle will not move.
• Not your stack. Yours should match your work, not mine. Take the principle first, pick the tool that fits second.

Close

Look at the five again. Every one is a post I wrote before the tool was on my radar. The method, the memory, the knowledge base, the tests, the design sense. I argued for the idea, then found the software that made it real, in that order. Never the reverse.

That is the only way to pick tools that last. The idea is the thing you keep. The tool is just where it lives this year. Take the principle, then go shopping.