Volo Volk | AI Engineer

AI engineer. I build constantly, at work and on my own. YouTube, tools, side projects.
When a tool gets hyped I try it on real work and post what I found.

[Volo Volk | AI Engineer @AlphaBOOLT]

@Joshua_WD True

[Volo Volk | AI Engineer @AlphaBOOLT]

Yesterday I said: burn Fable 5.2 limits on testing skills. Then I sat the fuck down and did it myself.

30 minutes. One task:
– audit every single skill on my machine
– count how many times I actually invoked it over the last 2 months
– deliver a verdict: keep, cut down, or nuke

Turns out, Claude couldn’t even see a couple of them. The descriptions were too fucking long, so only the titles made it into context. The skill exists, but it’s completely useless.
Agents were calling tools that don't exist.

First real run in the wild, and the whole shit would've crashed.
Purged half of them to hell, trimmed down the rest.

Audit your shit.

[Volo Volk | AI Engineer @AlphaBOOLT]

Anthropic let you reset your Fable 5.2 limits, and you're already generating your hundredth purple dashboard?

Stop. Fucking stop.

Spend them on your skills instead:

1. Test all of them. Whatever doesn't fire, delete that shit.
2. Improve what's left.
3. Build new ones from what you actually do in your sessions.

The trick: if you're digging through all your old sessions, send a Haiku fleet to read them first. Fable only writes the summary. Otherwise your limit burns out in one evening.

Limits got reset, not gifted.

Which skill of yours dies first?

[Volo Volk | AI Engineer @AlphaBOOLT]

The mistake: the smartest model does fucking everything. Plans AND writes every line.

OpenCode Go, $10:
– DeepSeek V4 Pro only plans ($15/mo limit)
– V4.1 Flash writes the code ($100/mo limit)

+$20 Claude Code: Opus plans, DeepSeek still does the grunt work

[Volo Volk | AI Engineer @AlphaBOOLT]

Everyone says you need a $100+ Claude/Codex plan to get into AI coding.

Bullshit. $10 is enough. Add $20 if you want premium.

But almost every beginner makes one dumb fucking mistake and burns the whole limit in a single evening 👇

[Volo Volk | AI Engineer @AlphaBOOLT]

@Lummox_eth Is Blender MCP in the room with us right now?

[Volo Volk | AI Engineer @AlphaBOOLT]

Your turn: do your skills actually do something, or just look smart?

Friday: drop your SKILL.md below. First 15 get a free public teardown:
- does Claude even see it
- how much fluff eats your context
- what nobody actually checks
- 3 fixes

Don't get butthurt, I'll be honest.

[Volo Volk | AI Engineer @AlphaBOOLT]

A skill called unslop is going viral. 10s timelapse: a purple vibecoded dashboard turns clean, same data.

Looks good? Yeah, fucking good. Right idea? Also yes.

But I can't just bookmark it and move on. Let's see what the fuck is actually inside. 🧵

[Volo Volk | AI Engineer @AlphaBOOLT]

What I changed:
- 21 slop signs caught by a script, not the model's eyes
- contrast, spacing grid, colors measured after every pass
- data numbers checked before vs after, none may change
- video renders only from measured numbers. No measurement, no video, no bullshit

[Volo Volk | AI Engineer @AlphaBOOLT]

Ladder in CLAUDE.md for beginners, hook for the rest.

What's the dumbest overengineering Claude pulled on you?

[Volo Volk | AI Engineer @AlphaBOOLT]

You ask Claude to fix one button.

You get 3 new files, a new library, a custom "state manager" and a wrapper around a wrapper.

"I've made the solution more robust and scalable."

Why the fuck do I need a scalable button? 🧵 #ClaudeCode

[Volo Volk | AI Engineer @AlphaBOOLT]

Too lazy to check the diff yourself? Stop hook.

When Claude tries to say "done", the script checks git for new files and dependency changes. Found any? It sends Claude back: justify each in one line or remove it. Second attempt passes, no infinite loop.

[Volo Volk | AI Engineer @AlphaBOOLT]

git diff --stat after every task:
– package.json or lockfile changed for 5 lines of logic? Someone else's code you now maintain. Revert
– interface with one implementation? Revert
– new file you can't justify in one line? Delete

10 seconds of looking.

[Volo Volk | AI Engineer @AlphaBOOLT]

Every extra line is one you'll read and debug later.

The model reads it too. Every time it opens those files, the junk goes into context and you pay for the tokens.

More code, more places to break.

[Volo Volk | AI Engineer @AlphaBOOLT]

The ladder is a prompt. And the model decides for itself whether it followed it. It always decides yes.

Writes a helper next to an identical one and reports "kept it minimal".

Don't trust the report. The report is its opinion. The diff is the work.