PLAYBOOK — the part where you actually build something
What this section is
The rest of the Tech Stack explains how the machine works — the chips, the data centers, the models, the frameworks that wire them together. This section is the other half: what to actually do on a Tuesday when you have a job to get done and a model to do it with.
It's the practitioner wing. Five pages, and they're arranged the way problems actually arrive — not the way a curriculum would order them. You start by prompting badly. Then you discover context is the real constraint. Then you hand work to an agent and learn what happens when it's wrong. Then you bolt on retrieval, try to prove any of it works, get the bill, and finally someone attacks it.
Nothing here is theory for its own sake. Every guide is something you can run today, and the failure modes described are ones we have watched happen.
The five pages
Tips, Tricks & Best Practices. Seven sections in the order problems show up: prompting, context engineering, working with agents, RAG, evaluation, cost, and security. Each one gives you the principles, the numbered moves, and a "mistakes everyone makes" block. Its own advice on how to read it is the right advice: skip to the section that's currently on fire.
The Prompt Library. Copy-ready prompts across six categories — writing and editing, coding, analysis, and the rest. Bracketed placeholders are the only parts you change. Every entry says what the prompt actually does and gives one harder variation for when the basic version stops being enough. There's a copy button on each one; the page is built to be used, not read.
The How-To Guides. Six builds, each runnable end to end: run a local model tonight with nothing leaving your laptop · build your first RAG app · set up Claude Code and work agentically · build a team-docs assistant · put AI inside a spreadsheet workflow · evaluate two models on your task in thirty minutes. Same shape every time — what you'll build, what you need first, numbered steps, a "you know it worked when" check, and where to go next.
Integrations. A model that can't touch anything is a very expensive autocomplete. This is the wiring: MCP — the protocol that became the standard, including the security part everyone skips — plus raw API patterns, streaming and structured output, the no-code automation layer, and the enterprise path. Integration is where most AI projects quietly succeed or die.
IDEs & Platforms. The six coding tools that define the market — Claude Code, Cursor, GitHub Copilot, Windsurf, JetBrains AI, Zed — plus the tier below them, and a straight answer on which to buy depending on who you are.
Which door to use
You have a specific task right now — writing, refactoring, summarizing. Start at the Prompt Library. Copy, paste, go. Come back for the rest later.
You want to build a thing. Start at the How-To Guides. Pick the build closest to your goal and run it end to end before you customize anything.
Something you built is behaving badly. Go to Best Practices and jump to the section that matches the symptom. Bad outputs are usually prompting or context, not the model.
It works alone but won't connect to anything real. That's Integrations — email, CRM, codebase, database, browser.
You're deciding what to buy or what to standardize your team on. That's IDEs & Platforms.
Where it breaks
The sticker price is an entry fee, not a total. Nearly every tool in this market now meters model usage underneath the subscription. The gap between the advertised monthly price and what a heavy agent user actually pays can be several multiples of it. Every pricing section in these pages flags the meter, because the meter is the part that surprises people.
The tools are converging, so comparisons rot fast. Every editor is growing an agent; every agent is growing an editor. Open protocols are dissolving the walls between them. Distinctions that are real today may be gone in two quarters — check the review date before you trust a comparison, including ours.
Agent failures are quiet. A wrong answer announces itself. An agent that edits the wrong file, or half-finishes a refactor, or silently drops a requirement, does not. Most of the security and evaluation advice in this section exists because of that asymmetry.
And "it worked in the demo" is not evidence. The only benchmark that matters is your own task, on your own data, measured. There's a how-to for exactly that, and it takes thirty minutes.
What we will and won't tell you
No tool in this section is paying for placement. Where a verdict is unflattering, it stays that way.
Every hard claim — pricing, adoption figures, protocol dates — carries a verification tag until an editor has confirmed it against a primary source. Unverified figures stay visibly flagged on the page rather than being quietly presented as fact. Where a number is contestable we publish the conservative end and show the range, and everything is date-stamped. This market moves monthly; if a figure looks stale, it probably is.
Where this connects
Models — which model family to pick before you start prompting it.
Foundation — what these systems actually are underneath, if the why matters to you.
Orchestration — the frameworks and plumbing the how-tos assemble.
Resources — the full tool and model directory, broader than the six platforms covered here.