Layer 1 — durable
Changes on a scale of years
The method. How to frame a prompt, how to control an image model, how to review machine-written code, how to price that work and how to test a tool before you pay for it.
A book by Prague Core
AI That Works
The no-hype field guide: which tool for which job, how to steer it so it does what you asked, and what it really costs. Decision tables instead of opinions — and next to every answer, how you tell that it actually worked.
- 136 pages across 15 chapters
- 60+ tools checked against the vendors’ own pricing
- Checks of facts, layout and consistency
In progress Text and typesetting finished 4 September 2026 · in English
What it is
Most AI advice is either a hype thread or a ninety-minute video that ends with “it depends.”
This book is the opposite. Every chapter answers three questions, in this order: which tool, how to drive it and what it costs. Then it gets out of your way.
One idea runs underneath it. AI tools aren’t magic and they aren’t toys — they’re power tools with no safety guard: enormous leverage, zero judgement. The judgement has to come from a person, at two points: the specification up front and the verification at the end. The middle part, the part everyone posts about, is the easy part.
Layer 2 — perishable
Changes on a scale of months
The names and numbers. Model versions, per-token prices, who currently leads which leaderboard. Marked [SEPT 2026] throughout, with a ten-minute protocol in chapter 15 for re-checking the lot.
People who only learn Layer 2 have to relearn everything twice a year.
What’s inside
Fifteen chapters. The book itself is in English, so the titles need no translating here.
- The Map How to think about AI tools, the ten laws, and the ten-minute test that says whether a tool is worth your money.
- Prompting That Actually Works The six-part frame, the iteration loop, and the failure modes behind almost every bad output.
- Choosing Your Daily Driver Chat and reasoning models compared by job, price and honesty. Which one to pay for.
- Documents, Decks & the Inbox The unglamorous work that fills most weeks, and where the most hours are actually recovered.
- Code The highest-return use of AI for most professions. Autocomplete versus agents, the review discipline, and the seven ways AI code goes wrong.
- Images Model-by-job table, prompt anatomy, control (references, inpainting, LoRA), text in image, and what you may sell.
- Video Text-to-video versus image-to-video, shot language, character consistency, and the edit pipeline that hides the seams.
- Voice & Audio Speech synthesis, cloning, dubbing, transcription, sound design — plus the consent rules that keep you out of court.
- Music Getting release-quality tracks out of generative tools, stems, mixing, and the licensing minefield.
- Writing That Doesn’t Sound Like AI Voice capture, structural editing, and the tells to delete.
- Research, Data & Analysis Deep research agents, spreadsheets, document analysis, and a verification protocol that catches hallucinations.
- Agents & Automation What an agent really is, MCP, when automation pays for itself, and the guardrails that stop it burning money.
- The Stacks Complete tool stacks at $0, $60 and $250 a month, plus twelve end-to-end workflows.
- Rights, Risk & Money Copyright, commercial licences, privacy, disclosure law, a team policy, and how to price AI-assisted work.
- Cheat Sheets One-page references, the prompt library, the glossary, and how to re-verify this whole book in ten minutes.
How it was checked
On a book about AI this is the one thing you can either do properly or fake — and nobody finds out for a year.
Against the source, not an article
Every price, tier and product name was checked against the vendor’s own pricing or documentation page, not taken from a secondary source. Deprecation dates come from the vendors’ own published pages.
Law from the records
Legal claims were checked against court records, the US Copyright Office and EU primary materials.
What could not be verified is not in it
And where a matter is contested — a live lawsuit, a shifting licence, a benchmark that flatters whoever published it — it is labelled contested rather than presented as settled.
Preparation includes verification and layout checks: fact-checks against live pricing and terms, a licensing audit, legal claims checked against primary sources, a technical pass over every constant, an editorial pass, print QA over every rendered page, and a regression check after each of them.
One defect came back in five consecutive rounds — a claim corrected in one chapter and left standing in another. Because of it the book’s build runs two automated checks that compare the paired passages on every edit.
What it looks like
Typeset for 176 × 250 mm. Tables, verification boxes, and a verdict at the end of every chapter.
How to get it
The text is written, checked and typeset — but the book is not published yet. It has no ISBN and no decision on how it will be distributed; until then there’s no price and no download link here. When that is settled, it will appear on this page.
We built this page the way we’d build yours. Want one like it for your product?
Or simply: info@praguecore.com