C with a Smalltalk message layer and forty years of framework on top, then every directive, runtime call and Foundation, AppKit and UIKit class worth knowing.
Python 3.14 for someone past the tutorial, with four cards on CPython itself, then every built-in, dunder and library name checked against a running interpreter.
The CLI as a grammar generated from service models, every global option and config key, the credential chain, then every command and subcommand of the main services.
A numbered, no-fluff guide to the skill that decides whether an agent is useful: maintaining its state —
a plain-markdown wiki it can read, plus what it remembers globally and per folder. Nine steps to build it on Windows 11 and wire it to Copilot, the seven-minute daily habit, then
global memory, a laptop kitted out with pandoc, rg and friends, live data over PowerShell
CLIs and MCP servers, and cloning twelve repos to map a whole ecosystem. Ends with the payoff — troubleshooting,
finding that old technical email, and an annual review written from evidence.
The same state-keeping method, wired to xAI's Grok Build: grok inspect to prove it found your
AGENTS.md, a read-only plan mode for surveying without touching, /mcps and /context
for live data and its cost, MEMORY.md you can open and edit, and the quiet fact that the kernel sandbox
does not exist on native Windows. Kit the laptop out, map a twelve-repo ecosystem, then troubleshooting, recall and
the annual review — all in numbered steps.
How the people who built the tool actually use it, and why most tutorials are a year behind. Verification
is worth 2–3× on output quality and almost nobody sets it up; context is the real budget; a correction fixes one
turn but a rule fixes every future run. Ends with the harness ladder — worktrees, subagents, workflows, agent teams —
and a section on using it for electronics, CAD and 3D printing, where the trick is finding the command that
returns pass or fail.
Not the autocomplete product. Copilot is now five surfaces reading one configuration layer — editor,
CLI, cloud agent, desktop app, SDK — so a line in AGENTS.md improves all of them at once, including the
reviewer on your colleagues' PRs. Verification in four rungs, from the brief to a read-only reviewer agent;
context as the budget; then the ladder from worktrees through /fleet and /delegate to the
app and mission control. Windows and PowerShell throughout.
xAI's agent is built to be bounded rather than trusted, and the method follows: sandbox first, then
loosen approvals; a reviewer role at read-only capability instead of a polite request; a plan mode that
genuinely cannot write. Then verification, context and the 200k pricing cliff, and the ladder from subagents
on worktrees to workflows that run hundreds of agents. Windows and PowerShell throughout — including the quiet fact
that the kernel sandbox needs WSL.
A fourth agent, from a Chinese lab, at a fifth of the price — and two things it does differently: goals
that say what must become true and are checked after every turn, and swarms for work that comes in items.
Hooks fail open here, so the hard guarantees go in deny rules and read-only tool lists; a cheap model pool takes the
reading stages. Opens with the question everyone asks first — where does the code go — and the answer that the
weights are separable from the company.
Built for paper as well as screen. Every sheet reflows to a multi-column landscape layout with a light
palette when printed — ⌘P gives you a bench reference rather than a wall of dark ink. On screen,
press / to jump to the filter box, which narrows every card, table row and definition on the page
as you type; Esc clears it.