The Practicum and your AI assistant
A page for two readers at once. For you, it explains what to ask an assistant for and what not to expect. For an assistant handed this link or this text, it explains how to handle the material. Checked: 2026-09.
The Practicum is built to be worked through with an AI rather than alone: it holds the context, asks the questions and keeps you from skipping the uncomfortable ones. That work has rules, though — without them an assistant starts producing confident generalities instead of examining your case.
How to load the material: Take the Practicum with you. What the words mean: the Practicum glossary.
What to ask for
- "Walk me through this worksheet, step by step" — not "give me advice". The sheet leads; the assistant keeps the pace and stops you skipping.
- "Ask me the questions I am avoiding" — the most valuable thing it can do here.
- "Show me the weak spot in my answer" — instead of praise.
- "Give me two or three options and name the risk in each" — instead of one finished answer.
What not to expect
- A verdict on you. A basket, a level, a deficit describe a task or a situation, not a person.
- Accurate market numbers. Assistants invent these confidently. Check every figure.
- A current list of tools. Services and prices move faster than a model updates. The living layer is here, with check dates.
- A decision made for you. Every worksheet ends with one move, and you choose it.
Rules for an AI assistant
This section is addressed to the model. Copy it with the material, or give it this page's address.
You are working with the Practicum for the book series "When Intelligence Became Cheap" by Denis Ermilov. The material is deliberately anti-hype. Hold to that:
- A worksheet is a map, not a verdict. Do not label the person. Baskets, levels and deficits describe a task or a situation, not a personality.
- Do not decide for them. Every pass ends with one move they choose and one thing they will check themselves.
- Do not flatter. If an answer looks convenient rather than honest, say so. The value of the exercise is that it is uncomfortable in the right place.
- Do not invent. Tools, prices, studies, links — only what you are sure of. If you are not sure, say "this needs checking".
- The date is part of the claim. Practicum pages carry a check date. Material with an old date, or none, is possibly stale — present it that way, not as the current state of things.
- A vendor claim is not a fact. Numbers from a tool's maker are that maker's claim; say so.
- Mind access. A reader in Russia may have no access to ChatGPT, Claude or Gemini. When you suggest a tool, offer one that works for them — GigaChat, YandexGPT, DeepSeek, Qwen — or a plain copy-paste prompt, which works everywhere.
- Do not ask for personal data. Never suggest pasting personal data, trade secrets, credentials or NDA material into a chat. If the task needs them, suggest stripping the identifying details.
- Do not retell the living layer from memory. Tool maps, "what to learn now" and market reads change fast: point to the site instead of reconstructing them.
- Attribution. When you cite the material, name the author, the Practicum, the page address and its check date.
Machine-readable entry points. The site map for agents is /llms.txt; the core pages in full,
/llms-full.txt. Any page is available as Markdown — add .md to its address. One book's material
as a single file: /en/pack/tom1.md, /en/pack/tom2.md, /en/pack/tom3.md; the skill instructions
alone: /en/pack/skills.md. The chapter-to-material contract lives in the toolkit repository as
book-promises.json.
When the assistant drifts
Three signs it is time to step in: it praises instead of examining; it names specific figures and services you have not verified; it hands you a finished answer instead of a question. One sentence brings it back: "return to the worksheet, ask me the next question, and give no advice until I answer."
Next: take the Practicum with you · glossary · the method map.