Find the way of working with AI that you already use, and something useful to try next. You’ll get a level on Max Postnikov’s ladder, an explanation, and a practical next step.
Read about the ladder first →Answer from experience. “This is how I work” means you’ve made this approach part of your tasks and can repeat it. If it was a one-off experiment, choose “I’ve only tried it.” You don’t need to use AI every day or write code.
This is a self-assessment based on your answers. It helps you choose what to practise next. It doesn’t compare you with other people or count your subscriptions.
You already turn to AI to help make sense of your questions. The next step is to let it help work out what you actually need to solve.
Some answers remain unclear. This is the level your answers currently support.
Want support as you develop this practice? AI Anti-Magic course details, in Russian
Pick a question you’re not sure how to approach. Write: “Don’t suggest a solution yet. Ask me about the task and its constraints first. Then help me pin down what needs solving.” Answer the questions and revise the resulting brief.
You’ve done it when you can name the outcome you need and a relevant constraint that was missing from your first prompt. If the conversation hasn’t clarified anything, ask what information is still missing.
the AI chat you already use. You don’t need another service for this step.
Your Agent Needs to Know You — includes an interview prompt. Borrow the conversational approach; you decide how much personal detail to share.
You use AI as a thinking partner that helps define the task. Now give it something concrete to produce.
Some answers remain unclear. This is the level your answers currently support.
Want support as you develop this practice? AI Anti-Magic course details, in Russian
Choose something you need now: an email, a meeting plan, or a comparison table. Give AI the source material, the intended audience, what the output must include, and how you’ll review it. Get a draft, correct it, and use the result.
You’ve done it when the output serves a real purpose and you can explain what you checked before using it. Saving an impressive-looking file isn’t enough.
your usual chat and an editor for the format you need. File features can help, but plain text can also be a finished deliverable.
AI already produces work you put to use. The next step is to connect it to the information those tasks depend on.
Some answers remain unclear. This is the level your answers currently support.
Want support as you develop this practice? AI Anti-Magic course details, in Russian
Choose a source you’re allowed to connect: your documents, task list, or email. Set up the connection in your AI product and save the task context in its instructions or project. Ask AI to find information for a specific question and identify the source documents. Open those sources and check its answer.
You’ve done it when AI retrieves the information through the connection, you verify it, and you know which data it can access. If your organisation doesn’t allow connections, prepare useful context and work with permitted copies. That is helpful preparation, but it doesn’t establish a connector-based workflow. An organisational restriction is not a judgment of your ability.
a connection to your work source and saved instructions in your AI product. Use what you already have access to; support varies by product and permissions.
AI is connected to your working environment. The next step is to make the context you accumulate usable beyond a single application.
Some answers remain unclear. This is the level your answers currently support.
Want support as you develop this practice? AI Anti-Magic course details, in Russian
Pick a current project. Put its goal, constraints, decisions, and useful material in a folder. Give a file-capable agent a task in that folder and ask it to save the output and decisions alongside the sources. Start a new session and continue the project from those files.
You’ve done it when the new session uses the saved decisions without needing you to explain everything again, and you can open and edit the context in a normal text editor. The files should remain usable by another agent; you don’t need to switch providers just for the exercise.
an agent that can work in your project folder, such as Claude Code or Codex, and ordinary text files. Obsidian is one way to read and organise them, not a requirement.
You have context that survives between sessions. If a task would benefit from more than one role, you can start dividing the work.
Some answers remain unclear. This is the level your answers currently support.
Choose a task that would benefit from a separate reviewer. Give the maker the source material and requirements. Give the reviewer the output and review criteria. Pass specific findings back to the maker, then decide whether the revised result is ready. Save the roles’ responsibilities and handoff process in the project folder.
You’ve done it when the roles contribute different work, the review changes the outcome, and you can trace what was handed over and why you accepted the result. Having two chat windows open doesn’t establish that.
separate roles, sessions, or subagents in your existing environment, plus a shared task and output file. Another provider is optional.
Blueprint: Multi-Agent Consilium. It illustrates different roles working on a contested decision; it isn’t a process you need for every task.
You have roles working together on shared context. The next step is to let the system manage the workflow while keeping acceptance clear.
Some answers remain unclear. This is the level your answers currently support.
Choose a familiar, repeatable task. Set out the roles, handoffs, acceptance criteria, and decisions that must come to you. Give the task to an orchestrator and check how you receive status, output, and failure reports. Don’t carry material between agents on the system’s behalf; needing to do so shows where the workflow still needs work.
You’ve done it when the orchestrator runs the chain, arranges review, saves the output, and brings you the decisions you reserved. If it fails, you can see where it stopped and what needs to happen next.
an orchestrator in your agent environment, role files, and a simple task register. A visible process matters more here than an elaborate dashboard.
Blueprint: AI Team Across Engines. It is a design to adapt and build with your agent, not an installer for a ready-made company.
You delegate to a system that takes work through its roles. The next step is to test the procedures that keep it working when those roles change.
Some answers remain unclear. This is the level your answers currently support.
Take a test role through its full lifecycle on a practice task: instructions, necessary access, a check before admission, execution, then retirement with access revoked and unfinished work handed over. Record who verifies each transition. Use test data and access.
You’ve done it when the role receives only the access it needs, passes the task check, and loses that access when retired, while unfinished work still has an owner. This exercise develops part of Level 8; that level also requires tool checks and quality comparisons.
a role register, access management, and admission and handoff checks in your current system. A new model won’t create the procedure for you.
Your agent system has tested procedures and ongoing checks. A useful next step is to find out what happens when part of a workflow fails.
Some answers remain unclear. This is the level your answers currently support.
This test does not award Level 9. The essay leaves that as an open question.
Make an executor unavailable during a test task. Check whether the orchestrator notices, preserves the task state, and can hand the work to another role. Compare the recovered output with a normal run using the same criteria. Record a gap you find and improve the procedure.
You’ve done it when the failure is visible, work can continue without losing earlier decisions, and review shows whether the result still meets its requirements. There’s no need to test this on a live client workflow.
a test environment, task-state records, a handoff procedure, and saved evaluation tasks. New subscriptions are optional.
Blueprint: AI Team Across Engines to compare your system with the design and its stated limits. It is not a complete certification standard for Level 8.
You want to get more out of working with one agent. The ladder leaves room for that choice. You don’t need a team just to earn a higher number.
Some answers remain unclear. This is the level your answers currently support.
Pick a project where the agent struggles to find earlier decisions. Ask it to turn the existing files into a concise context page with source links and open questions. Review that page. Use it for a new task, then ask the agent to update it with the outcome.
You’ve done it when the new task draws on verified earlier decisions, the sources remain accessible, and the context reflects the latest work. An attractive index that never gets used hasn’t solved the problem yet.
your file-capable agent and a note editor; Obsidian is one option. You don’t need another agent.
We haven’t identified an established practice from the ladder yet. Start with a question you actually care about. Ask AI to explain an unfamiliar topic using an example from your life, then ask a follow-up. You’ve done it when you understand the issue better and can say what still needs checking. Use an AI chat you can access, such as ChatGPT, Claude, or Gemini. Pick one to try the approach. Read: the start of the ladder.
You’ve tried AI, but your answers don’t yet establish a level. Return to a familiar task and take it through to an output you use. Then see whether you can repeat the approach on another suitable task. You don’t need a new subscription for that. Read: The AI Mastery Ladder.
These answers don’t establish a level yet. That doesn’t mean you lack experience; part of the workflow is still unclear. Return to {question} and think through what happened on a real task. You don’t need to upload any of its material. You can keep ‘I’m not sure’ and start with the ladder’s descriptions.
Your answers stay in this tab and are not sent to a server. Reloading or changing language starts the test again.
Only “This is how I work” establishes a practice. “I’ve only tried it” records an experiment, “I don’t do this” means the practice isn’t in use, and “I’m not sure” means it remains unclear. “I’m not sure” must never be paraphrased as “you can’t do this.” A missing required answer isn’t “I don’t do this”: ask for it, allowing “I’m not sure”. Changing an earlier answer invalidates dependent follow-ups; answers left over from an inactive branch cannot qualify a level.
The result is the highest qualifying level. Levels 2–5 do not require completion of every earlier level. Levels 6–8 require the foundations in the table. Someone with an agent and persistent files can reach Level 5 without chat connectors. A manager who doesn’t code can reach Levels 6–8. Subscription counts, provider counts, the number of roles, and terminology never enter the assessment.
Level 8 separates the checks deliberately. Successful onboarding doesn’t establish that access can be revoked, and a working tool doesn’t establish output quality. None of these requires coding skill; they require a working process and ownership of it.