$ cat ./mind-palace/my-llm-toolkit.mdx

my llm stack: the tools i actually use when no one's watching

·4 min read·#ai·#tools·#llms· all posts

After drowning in a sea of models, tools, and “game-changing” AI platforms that absolutely weren’t, I’ve finally found my sweet spot. Here’s the stack that actually works without making me want to throw my laptop out the window.

The Harnesses (Where the Magic Happens)

Codex CLI is my main workhorse. GPT 5.4 and 5.5 live here, and this is where most of my actual workflows run. It’s the agent harness I reach for when I need something done without drama.

Opencode CLI is my sandbox. I use it to test small, temporarily-free open-source models and to tap my z.ai GLM subscription. It’s the “let’s see if this cheap thing works” harness. Sometimes it surprises me. Often it doesn’t. That’s the game.

T3 Code is the harness for my harnesses. Look, TUIs are great until you paste an image and realize you’re staring into the void because the model can’t see a damn thing. T3 Code gives me actual visibility over my history, and I can see the images I paste. It’s open source, way less laggy than the Codex app, and honestly? It just clicks with me better than anything else in this category. (Sorry, Jean.)

[Custom: Terminal-comparison-screenshot] [R: Illustrates the TUI vs GUI difference so readers understand why someone would layer a harness on top of another harness] [V: A split-screen mockup. Left side: a dark TUI terminal with a broken image icon and question marks where a screenshot should be. Right side: T3 Code’s interface showing the same conversation with the image rendered clearly in the chat history. Minimal UI chrome, dark mode]

The Skills (The Secret Sauce)

I don’t prompt from scratch like a caveman. My main driver skills are all from skills.sh.

The first four are Matt Pocock’s, and yes, they’re as good as you’d expect:

  • grill-me
  • to-issues
  • to-prd
  • improve-codebase-architecture

The rest are Anthropic and Vercel staples:

  • frontend-design (Anthropic’s)
  • vercel-composition-patterns
  • vercel-react-best-practices
  • web-design-guidelines (the Vercel trio)

I’ve also built a few custom skills to speed up my personal flow, but those are mostly micro-optimizations. I’ll write about them separately when I have the energy.

Skills inventory

Filter the stack by who made what.

grill-me

Matt Pocock

Pressure-tests a plan through a focused interview.

to-issues

Matt Pocock

Turns a plan into implementation-ready GitHub issues.

to-prd

Matt Pocock

Synthesizes context into a PRD.

improve-codebase-architecture

Matt Pocock

Finds refactors that improve boundaries and testability.

frontend-design

Anthropic

Builds polished, production-grade web interfaces.

vercel-composition-patterns

Vercel

Guides scalable React component composition.

vercel-react-best-practices

Vercel

Keeps React code aligned with current platform patterns.

web-design-guidelines

Vercel

Reviews web UI quality, accessibility, and interaction polish.

Custom Tooling (Vibe-Coded, Obviously)

When the ecosystem doesn’t give you what you need, you build it. Or rather, you vibe-code it and pretend you architected it.

Skill Manager — A tiny app to toggle skills globally or per-project. Because manually editing config files every time you switch contexts is a special kind of hell, and I refuse to live there.

Codex Limit Indicator — A macOS status bar widget that tracks my usage limits across multiple Codex accounts. Yes, I have multiple accounts. No, I won’t explain why. Yes, I need to know when I’m about to hit a wall.

[Custom: Menu-bar-mockup] [R: Shows the utility of the limit indicator in a format people instantly recognize—a macOS menu bar] [V: A zoomed-in macOS menu bar showing a small icon with “Codex: 12/30” next to the WiFi and battery icons. Clean, native-looking, no custom window chrome. Maybe a subtle green/yellow/red dot for usage status]

The Misc Drawer (Search, Research, and Vibes)

Perplexity is my search default right now because I have a free year subscription. When that honeymoon ends, I’ll probably move to a mix of Google + Gemini. For now, it works.

I route everything through unduckling so I can keep using DuckDuckGo-style bangs (!gh, !so, etc.) but with Google as the actual search engine. It’s one of those “why isn’t this built-in” tools.

For deep research, ChatGPT 5.5 Pro is genuinely insane. But when I need something faster and less… thorough, Perplexity gets the nod.

And for anything I actually write—like this post—I reach for Kimi. It’s the model whose stylistic vibe matches mine the closest. It just sounds like me, or at least like a version of me that proofreads.

[Custom: Search-routing-diagram] [R: Clarifies the unduckling setup because “bangs with Google backend” is slightly cursed and needs visual explanation] [V: A simple flow: User types query with !gh → unduckling intercepts → routes to Google with site:github.com modifier → results displayed. Three boxes, two arrows, done]

That’s the stack. No AI-native IDEs, no “second brain” apps that require more maintenance than my actual brain, and definitely no tools that demo well but collapse under real work. Just a few harnesses, some well-curated skills, two vibe-coded utilities, and a search setup that doesn’t make me angry.

It’ll probably change in three months. That’s the hobby.