AI Tools & Skills
A working guide to the AI stack I actually run. An agent is its model plus its harness — and almost all of this page is harness.
AI Tools & Skills
A working guide to the AI stack I actually run. An agent is its model plus its harness — and almost all of this page is harness.
An agent = its model + its harness
On its own, a model takes text in and gives text back. Reading a file, running a command, remembering the last turn: all of that is the harness. It assembles the context for each request to the provider, runs the tool calls the model asks for, feeds the results back in, keeps the session history, asks me before anything risky, and decides when to compact. The loop is the harness’s too — the model proposes, the harness executes, and round it goes again.
01 / 10
What do I need to get started?
Two things, before any of the tools below: an account with a lab, and a machine that can run a command line.
Accounts
A handful of sign-ups unlock everything below. Most are free to try, and the paid ones bill either monthly or by the token.
GitHub
Where code and Agent Skills live. Make this account first — you will clone repositories, keep your own work, and install most of the tools on this page from here.
Free
Anthropic
The maker of Claude and Claude Code. A Pro or Max subscription powers the coding agent; an API key covers anything billed by the token.
Claude Pro from $20/mo · API pay-as-you-go
OpenAI
The maker of ChatGPT and Codex. The same plan that runs the chat app drives the coding agent and the agent on my server.
ChatGPT Plus from $20/mo · API pay-as-you-go
Google AI Studio
A free Gemini API key in a couple of clicks. Handy for image work and quick model tests.
Free tier
Figma
Design files, and the account behind Figma Make and the Figma MCP server further down this page.
Free tier
Hardware
Command-line coding wants a laptop. How much laptop depends entirely on where the model does its thinking, and the answer is smaller than most people assume.
Frontier models on a plan
The compute happens in someone else’s data centre, so your machine is mostly a terminal with a good keyboard. 16GB of RAM does it.
16GB RAM
Small models, run locally
Running an SLM on the machine itself is the case that needs real memory, because the model has to sit in it. Plan for at least 24GB, and 32GB or more if you can stretch.
24GB minimum · 32GB and up
A fan
Local inference runs hot for long stretches, and a fanless machine throttles under it. My 32GB MacBook Air overheated running small models, which is what moved me to a MacBook Pro.
Only if you run SLMs
02 / 10
Models & providers
Five providers, and I use all of them. Which one I reach for depends on how much the answer matters and what it costs to get it.
Providers
The labs whose models do the actual thinking. Everything in the sections below is a way of driving one of these.
Anthropic
Claude, and Claude Code, which is where most of my working day goes.
Claude Pro from $20/mo · API pay-as-you-go
OpenAI
GPT and Codex. One subscription covers the chat app, the coding agent and the autonomous agent running on my server, which is the reason it earns its place twice on this page.
ChatGPT Plus from $20/mo · API pay-as-you-go
Gemini, including the image models. The free API key from AI Studio goes a long way before you ever pay.
Free tier · API pay-as-you-go
xAI
Grok, which I reach through its own command line rather than through a browser.
Signed in with an X account
Ollama
Runs small models on the Mac itself. With enough memory, routine work stays local and off the token meter.
Open source · Free
Model comparison
Where I check which model to reach for. Both earn a picture, because the chart is the answer.
Benchmarks
Artificial Analysis
Free
Independent benchmarks that compare models and providers across intelligence, speed and price, so you can pick the right one for a job.
Leaderboard
Arena (LMArena)
Free
A crowd-sourced leaderboard that ranks models by blind, side-by-side human votes — millions of them, scored with Elo ratings.
03 / 10
Agents & tools
The shell you sit in front of all day — the coding agents, the terminal that holds them, the apps — plus the connectors that hand an agent a real tool. It reads your code, runs commands and carries a task through to a finished change.
Coding agents
Three of them are installed here, and two of them do the work.
Coding agent
Claude Code
Pro $20/mo · Max $100 or $200/mo — for real work you will want at least Max
Anthropic’s agentic coding tool — terminal, IDE or web, carrying a task from request to a shipped change. It is the stronger coder for me. One catch: the subscription only powers Claude Code itself. To drive other agents you drop down to the pay-per-token API, which adds up fast.
Coding agent
Codex
Plus $20/mo · Pro $100 or $200/mo — the $100 Pro tier is enough to run agents on your plan
OpenAI’s coding agent for terminal, IDE and web. Its quiet advantage: the same ChatGPT plan can power other agents too — the autonomous agent on my server runs on this subscription, where Claude Code would bill me separately through the API. On sharing a plan, Codex wins.
Grok CLI
xAI’s agent on the command line, signed in through an X account rather than an API key.
Included with an X subscription
The terminal
Where the agents actually run, and how you keep a fleet of them in view.
Terminal
cmux
Open source · Free
A native macOS terminal built on Ghostty, with vertical tabs, split panes and notification rings — built for running several AI coding agents side by side. It is the piece of this whole stack I touch most.
cmux setup guide
Running agents in parallel
Free
Artem Zhutov’s guide to driving a fleet of agents from cmux — an orchestrator on top, with a dashboard, sessions and workspaces underneath, so ten tabs stop becoming ten guesses.
Desktop apps
The same models in a window instead of a terminal. Both are installed here, and both carry the same limit: swapping models in and out is awkward in a way a command line makes trivial.
Claude
Anthropic’s desktop app. Good for thinking a problem through before any of it reaches a repository.
Free · Pro from $20/mo
ChatGPT
OpenAI’s desktop app, and where the voice work in section nine happens.
Free · Plus from $20/mo
Menu-bar tools
Small native apps that keep an agent fleet in view.
Menu bar
CodexBar
Open source · Free
Usage limits, credit balances and reset countdowns for Codex, Claude Code and 40+ AI coding providers, all from the macOS menu bar.
RepoBar
Keeps GitHub work in view from the menu bar — CI status, open issues, pull requests, releases and rate-limit health for every repo you follow.
Open source · Free
MCP servers
The connectors that hand an agent a real tool. They plug into the harness rather than into the model, which is why they sit here. Two more live with their own tools: the GitHub server in the next section, the Figma server in section nine.
qmd
Local search across my markdown vaults — keyword, vector and reranked together, so an agent can find the right note without me naming the file.
Runs locally
Claude in Chrome
Drives a real browser tab — clicking, filling forms, reading the console — so an agent can check its own work in the thing a reader will actually use.
Included with Claude
04 / 10
Context & knowledge
How I feed and maintain the context the agents work from.
The vault
One store, in plain files, that everything else reads from and writes back to.
Knowledge base
Obsidian
Free for personal use
A local-first markdown knowledge base. Your notes stay as plain files on your own machine, linked into a graph you control — the store everything else on this page feeds.
Plugins & knowledge bases
The add-ons I run on top of the vault, plus the pattern worth pointing an agent at. Two of them change what a folder of notes can do, and both are easier to see than to describe.
Obsidian plugin
Dataview
Open source
Turns your notes into a queryable database — list, table and task views generated live from frontmatter and inline fields.
Obsidian core
Graph
Free with Obsidian
Every note a node, every link an edge, so the shape of what you know is something you can see.
Web Clipper
A browser extension that saves any web page as clean markdown straight into your vault, with templates for the metadata you want.
Sync
Keeps my vaults in step from the server down to the Mac and across every device, end-to-end encrypted, with version history I can roll back.
$8/mo
Excalidraw
Hand-drawn diagrams and sketches you draw straight inside a note, kept alongside the writing.
Open source
Bases
Obsidian’s built-in way to turn a folder of notes into a database view — filter, sort and group without leaving the vault.
Free with Obsidian
LLM wiki
Andrej Karpathy’s pattern for AI-maintained knowledge bases: the model builds a cross-linked markdown wiki that compounds as sources and questions accumulate.
Marker
Turns PDFs, slides and documents into clean markdown the agents can read — tables, structure and all. It is how anything that arrives as a document gets into the vault.
Open source
05 / 10
GitHub
Skills, plugins and most of the tools on this page install from GitHub, so it is worth getting comfortable with it before the next section.
How I use it
Beyond the account in section one, three things earn their keep.
gh CLI
GitHub from the command line. The agents use it for issues, pull requests, checks and releases without ever opening a browser.
Open source · Free
GitHub MCP server
The same surface again, this time as a tool the agent calls directly — useful when it needs to read a repository it has not cloned.
Open source · Free
Worktrees
Several agents on one repository without them colliding: each gets its own checkout on its own branch, so nobody is switching branches under anyone else. Non-obvious, and it changes how many agents you can usefully run at once.
Built into git
06 / 10
Skills
A skill is a folder of instructions an agent loads when the work matches it. Anthropic ships a handful to start with, and most of the rest come off GitHub.
Anthropic skills
The starters I reach for most — document work and builders.
Fills forms, merges and splits files, and pulls text and tables out of PDFs — the document work that used to mean a dedicated tool.
Open source
docx
Creates and edits Word documents with tracked changes, comments and formatting kept intact.
Open source
xlsx
Builds and edits Excel workbooks — formulas, formatting and multiple sheets — and reads the numbers back.
Open source
pptx
Generates and edits PowerPoint decks, keeping layouts, themes and speaker notes in place.
Open source
skill-creator
Authors new skills — scaffolding the folder, instructions and examples that teach an agent a repeatable task.
Open source
mcp-builder
Generates MCP servers, wiring a tool or an API into something an agent can call directly.
Open source
Claude Code commands
The built-in commands, and what I use each one for. The first is the one that changes how the work goes.
Built-in command
/plan
Built in
Researches the codebase and lays out an approach before changing anything. You approve the plan, then it executes. Toggle it with Shift+Tab.
/goal
Set a finish line and it keeps working across turns, checking after each one, until the condition is met.
Built in
/workflows
Watch and manage a workflow — a script that runs many subagents in parallel — as it works in the background.
Built in
/loop
Runs a prompt on a repeating interval, or paces its own iterations when you leave the timing to it. Section eight is what I point it at.
Built in
/agents
Defines subagents you can hand specialised work to, each with its own tools and instructions.
Built in
/rewind
Rolls the code and the conversation back to an earlier checkpoint when a path turns out wrong.
Built in
Notable skills
Individual skills worth borrowing. Artem Zhutov draws his, and the drawing is the argument, so his three keep their diagrams.
Claude Code skill
retrospective
Free
Artem Zhutov’s /retro. At the end of a session the agent looks back, works out what it learned and writes it down for next time. Skip it and every session starts from zero; run it and the learning compounds across everything you do.
Claude Code skill
handoff
Free
Artem Zhutov’s /hand-off. A context window fills up and the agent dulls in the last fifth. This packages the live state — the work, the session, where you are — and passes it clean to a fresh agent, so a long task survives the boundary.
Claude Code skill
skill-management
Free
Artem Zhutov’s approach to growing a skill. Start with everything in one SKILL.md. When that file gets too big, split the parts into workflows and let the main file route to them, so the agent loads only the workflow it needs.
gstack
Garry Tan’s opinionated Claude Code setup — 23 tools that act as CEO, designer, engineering manager, release manager and QA inside your agent.
Open source
last30days
Matt Van Horn’s research skill. Pulls posts and engagement from Reddit, X, YouTube, TikTok, Hacker News and Polymarket over the last 30 days, then synthesises a grounded, cited summary of what people are actually saying.
Open source
Where skills come from
Most open skills ship as a GitHub repository — a folder of instructions, scripts and examples you can read, fork and drop into your own agent. These are where I look.
Marketplace
ClawHub
Free
A skill marketplace for Claude Code and OpenClaw with vector search — the closest thing yet to npm for AI agents.
Peter Steinberger
Ex-PSPDFKit founder, now at OpenAI and steward of OpenClaw. His GitHub is the source behind CodexBar and ClawHub — two of the tools already on this page — plus agent-scripts, a shared rule set for coding with agents.
Free · GitHub
Awesome Evals
A curated library for building and evaluating AI agents — papers, talks, tools and benchmarks, plus reading notes on eval frameworks, LLM-as-judge and agent-trajectory testing.
Free · GitHub
Awesome MCP Servers
A large community list of MCP servers — the connectors that hand agents real tools, sorted by what they do.
Free · GitHub
Awesome Claude Connectors
A directory of Claude’s connectors — hundreds of verified MCP integrations sorted into 30+ categories with descriptions and use cases.
Free · GitHub
07 / 10
Plugins & the engineering process
A plugin does for a whole setup what a skill does for one task. Two of them matter here: a process that shapes how the work gets done, and a design critic I invoke more than anything else I have installed.
From every.to
The process
Every piece of work should leave the next one easier.
Claude Code plugin
Compound engineering
Open source · Free
Kieran Klaassen’s plugin for Claude Code, built on a single idea from Every: every piece of work should leave the next one easier. Solved problems get written down as you go, so the setup compounds the more you use it. It ships 26 agents, 23 commands and 13 skills.
The loop
Brainstorm → Plan → Work → Review → Compound, then repeat. Compound is the step most AI coding setups skip.
Brainstorm
Turns a fuzzy idea into a sharp, agreed requirement before any code is written.
Plan
Researches the codebase and writes an implementation blueprint to work from.
Work
Executes the plan in an isolated worktree, validating as it goes.
Review
Runs a panel of specialist reviewers in parallel — security, performance, architecture and more.
Compound
Writes the solved problem into docs/solutions/, so the next run starts ahead.
LFG
The end-to-end pipeline: one command takes an idea all the way to a pull request.
Reviewers and skills
Behind the loop sit 14 specialist review agents — security, performance, architecture, data integrity, code quality and framework-specific reviewers for Rails, Python and TypeScript — plus 13 skills holding the architecture patterns, style guides and design systems the agents read while they work.
Plugins I run
Installed from marketplaces. These four still get used months later.
impeccable
A front-end design plugin — direction, critique and polish for anything with a viewport, judged against the reference at desktop and mobile widths. The most-invoked thing in my whole setup, by a distance.
Open source · Free
codex
Runs Codex from inside Claude Code, so one agent can hand a stuck job to the other and read the answer back. Second by usage in my setup, and the value is in the pair.
Open source · Free
mattpocock-skills
Matt Pocock’s engineering skills — domain modelling, TDD, codebase design, and the two grill skills below.
Open source
hands-on-deck
PowerPoint, properly: reading a deck, editing text and images in place, building new slides and checking the result visually.
Open source
Stress-testing the plan
Before the loop runs, the plan gets interrogated. Two of Matt Pocock’s skills built for exactly that.
grill-with-docs
The grill interview run with the project docs loaded — it questions the plan in your domain language, against your ADRs and terminology, and writes the decisions back into the docs inline.
Open source
grill-me
Interviews you about a plan until every branch of the decision tree is resolved.
Open source
08 / 10
Loops & hosted agents
Agents that run on their own — on a loop, or on a server that never sleeps.
Loops & autonomy
Point an agent at a metric or a schedule and it keeps going — optimising, repeating, iterating — until the job is done or you call it off.
Loop library
Loop Library
Free
Forward Future’s collection of practical agent loops — reusable prompts with clear success checks and stopping conditions, across engineering, evaluation, operations, content and design. Copy a loop straight in, or install the Loopy skill to run them from your agent.
looper
Design and scaffold an agent loop: set the scope, the metric and the keep-or-revert rule, and it writes the harness for you.
Open source · Free
autoresearch
Andrej Karpathy’s loop for autonomous ML research: the agent edits training code, runs short timed experiments and keeps only the changes that move the metric.
Open source
Hosted agents
The always-on side: an autonomous agent on a rented server, working while the Macs are asleep.
Hosted agent
Hermes
Open source · runs on my Codex subscription
Nous Research’s autonomous agent. Mine keeps the vaults in order, fetches market data, writes the newsletter and the tweet digests, maintains a signal ledger, runs the investment analysis and sweeps its own cron jobs for quality — all on a schedule, without me. It runs on Codex, which is the whole reason a second subscription was never needed.
Hosted server
Where the agent lives. It has to be somewhere always on, and I do not keep a Mac running around the clock, so I rent a box. Aim for 8GB of RAM and 80GB of disk — enough headroom for the agent and its state.
Hostinger VPS
09 / 10
Voice, image & making things
The specialist tools: pictures, voice, and the things I make with them.
Images
Everything visual on this site starts in one of these two.
Image generation
ChatGPT Images 2.0
Image generation in ChatGPT · Plus from $20/mo
My favourite image generator. It follows a detailed brief closely and lets you edit a picture conversationally until it is right. The image work on this site runs through it.
Google AI Studio
Google’s playground for Gemini, including the Nano Banana image model — strong at editing a picture and holding a subject steady across generations.
Free tier
Voice
A lot of my first drafting happens out loud.
Dictation
FluidVoice
Open source · Free
A free, open-source Mac dictation app. Its on-device Fluid-1 model cleans spoken words into properly punctuated text in any app, across 40+ languages, with nothing leaving the machine.
ChatGPT voice
The other half of it: talking a problem through in the ChatGPT app on a walk, then bringing the answer back to a keyboard.
Included with ChatGPT
Design & making
Where a design becomes a file, and a file becomes something an agent can build from.
Design
Figma
Free tier · Pro from $16/seat/mo
Design files, plus Figma Make, which builds a working prototype from a description and lets you edit it by hand. Its MCP server is the part an agent uses: it reads a file’s structure, variables and components, and turns a frame into code.
Claude Code plugin
Frontend Slides
Open source · Free
A Claude Code plugin that builds animation-rich HTML slide decks from a prompt or an existing PowerPoint. Zero dependencies, a library of bold templates and a design philosophy aimed squarely at avoiding AI slop.
Typography
Tufte CSS
Open source · Free
Tools for styling web articles after Edward Tufte’s books — generous sidenotes, tight text-and-figure integration and careful typography.
Tufte data viz
Caylent’s skill, applying the same principles to charts across Recharts, ECharts, Chart.js, matplotlib, Plotly and D3.
Open source
Local tools
Smaller things I reach for around the agents.
Playwright
Drives a real browser from code — clicking, typing, screenshotting. It is how the agents capture pages and check their own work.
Open source · Free
Typora
Where Obsidian holds the linked vault, Typora is for the single document. A clean, distraction-free live preview for drafting one piece and exporting it to PDF, Word or HTML.
$14.99 one-time · 15-day trial
birdclaw
Steipete’s local-first tool for your X account. It pulls your tweets, DMs, likes and bookmarks into a database you own, with a web view to search, read and post.
Open source
10 / 10
People
Who I learn from. Several of the tools above came off these timelines before they reached anywhere else.
People I follow on X
The accounts I learn the most from. Several of the tools above came off these timelines first.
Peter Steinberger
Ex-PSPDFKit founder, now at OpenAI and steward of OpenClaw. CodexBar, ClawHub and birdclaw all come off his GitHub.
Artem Zhutov
The retrospective, handoff and skill-management skills on this page are his, and so is the guide to running a fleet of agents in cmux.
Andrej Karpathy
The LLM wiki pattern in section four and the autoresearch loop in section eight are both his.
Garry Tan
gstack, his opinionated Claude Code setup, is in section six.
Matt Van Horn
The last30days research skill in section six.