What I'm Stealing From The Claude Code Tools Video
Watched a 40-minute tour of every shiny Claude Code add-on. Here's the short list I'm actually installing — and the mental model behind the cuts.
Source: youtu.be/KjEFy5wjFQg · Adapted 2026-04-10
The Short Version
Imagine your brain is a kitchen. Every time someone releases a new kitchen gadget, you could cram it into a drawer. Some will save you real time — a sharper knife, a timer you can actually hear. Others will sit there forever, smug and shiny, taking up room you wish you had for flour.
New Claude Code tools come out every week. The temptation is to install them all, because they're free and they look cool in a demo. But a drawer full of gadgets you never use is worse than no gadgets — it slows you down every time you go looking for the knife.
So the rule I landed on is simple: only keep a tool if it patches a place where my workflow actually bleeds. If I never cut myself, I don't need a fancier bandage.
The brutal part: most tools are bandages for bleeding that isn't happening. You just didn't notice, because the demo made the wound look real.
The Cuts I'm Making
The video walked through dozens of tools. I'm writing this to force myself to commit: here's the short list, and for each one, the specific leak it's plugging in my loop. Everything else gets left in the drawer.
.md. Claude can already grep it. No parsers, no APIs, no lock-in. Already how I work — just doubling down.DESIGN.md into a project and agents generate brand-matched UI. Already rewired kira-learn and Mission Control this way. The markdown-file-as-design-system move is underrated./notebooklm CLI — underused. Commit to running one a week on active research.What I'm Deliberately Not Installing
Google Workspace CLI sounds useful and is a hat for me right now — Marline already handles calendar and mail in the loop where they matter. Adding a second path would just be a place to forget updates. Same with multi-model orchestrators: interesting in demos, but I only have one use case (Claude), and routing between three models optimizes a problem I don't have.
The best tools are the ones that reduce weak spots in the loop. Everything else is plugin vanity.
The Filter Question
Before installing anything new, the question is no longer "is this cool?" It's:
- Which step of my loop is this patching? — name it out loud. Write, review, ship, research, docs. If you can't name the step, the tool doesn't get installed.
- What does the leak look like today? — a bug I missed, a context I lost, a test I skipped. If you can't describe the leak, there isn't one.
- Can I measure the before and after? — even crudely. "I caught 3 regressions last week vs 0 before." If not, I'm buying vibes.
The Audit I Actually Ran
This framing isn't just for new tools. I ran the same three questions against every Claude Code skill and MCP server I already have installed. Uninstalled four of them. Nothing broke. That's the test — if removing a tool doesn't break anything, it wasn't earning its drawer space.
# For each installed skill / MCP server:
# 1. Which loop step does this patch?
# → If you can't name it in 3 seconds: uninstall
#
# 2. When did I last use it on real work?
# → "Can't remember" means no.
#
# 3. If I removed it right now, what breaks?
# → "Nothing really" → gone.
#
# Run this monthly. The drawer fills up faster than you think.
Why Markdown-First Keeps Winning
One principle the video reinforced and that I keep seeing play out: the knowledge structures that pay off long-term are plain markdown. Design systems, project briefs, research notes, meeting summaries — all .md, all greppable, all Claude-native. The moment a tool demands its own format, you're adding friction that compounds for years. Markdown is the one format that never becomes legacy.
Key Takeaways
- Patch leaks, don't collect gadgets. A tool only earns its spot if you can name the specific step of your loop it's fixing.
- Adversarial review is the highest-leverage install. It catches the class of bugs that human + AI collaboration is worst at.
- Markdown-first is a compounding bet. Vaults, design systems, briefs — plain
.mdnever becomes legacy. - Benchmark, don't vibe. If you can't measure a skill's output against test cases, you're trusting it for the wrong reasons.
- Audit monthly. Ask what breaks if you remove each tool. Usually nothing. That's your answer.