Brad Zhang / public archive

Brad Zhang

AI product notes, agent workflow writing, and open-source dossiers for founders and early technical teams.

Back to X dispatches

Article / April 29, 2026

English indexed dispatch

Codex can actually not only do coding, but can also be used as a training partner! Try...

Codex can actually not only do coding, but can also be used as a training partner! Try the following: 1. First define "real mastery". Don't regard "understanding" as mastery. Fo...

Codex can actually not only do coding, but can also be used as a training partner! Try the following:

1. First define "real mastery". Don't regard "understanding" as mastery. For a skill to truly enter your mind, it must meet at least 5 criteria:

1. You can explain its core model clearly without using materials.

2. You can complete a real small project independently.

3. You can explain why you do it instead of just copying it.

4. You know where to look when a bug comes up.

5. In another scenario, you can still migrate and use it. So the goal of Codex is not to “give you knowledge” but to continually verify these 5 things.

2. Change Codex to training mode. You can give Codex a fixed instruction: From now on, you are not my assistant to complete tasks for me, but my skills training coach. Rules:

1. Don’t give full answers unless I specifically ask for it.

2. Ask me about my current understanding first, and then point out blind spots.

3. Let me repeat each knowledge point in my own words.

4. Design a small exercise for each study session.

5. After I submit the answer, you have to point out the problem like a code review.

6. Finally generate the wrong question book, review cards and next training tasks. The goal is for me to really grasp it, rather than rush through it.

3. A closed loop of Codex learning for a skill. It is recommended to use this loop: Concept input → Self-explanation → Codex correction → Small project practice → Review → Wrong question book → Interval review → Transfer challenge. Let Codex play different roles in each round: - Teacher: Explain the minimum necessary knowledge clearly. - Interviewer: Ask you why. - Reviewer: Review your code/solution. - Examiner: No prompts, let you do it independently. - Accident review officer: analyze where you went wrong. - Training Planner: Schedule your next practice session.

4. The most effective prompt word template When learning a technology, you can start like this: I want to truly master {skill/technology}. Please do not impart knowledge directly. Please train me as follows:

1. Test my current level with 5 questions first.

2. Based on my answers, give me a competency map.

3. Identify the most critical 20% of basic concepts.

4. Let me explain each concept first, and then let you correct it.

5. Design 3 progressive practical tasks for me.

6. Do a code review/design review after each task is completed.

7. Maintain a wrong question book and review plan.

5. The key is "active recall", not reading. You should always let Codex ask you, rather than you asking Codex. For example: Test me now without prompting. Ask me 10 questions about what you just learned. Just ask one question at a time. After I answer, you can judge: - Whether it is correct - Where it is vague - How it should be strengthened. This is much more effective than "summarizing xxx".

6. Use projects to push knowledge into experience. To truly form experience, you must do projects, and the projects must have constraints. For example, when learning Agent engineering, don't say "teach me Agent". Instead: Design a 7-day training program for me to implement a working Agent harness from scratch. Requirements:

1. There are runnable products every day.

2. There are engineering problems that must be stepped on every day.

3. You only do reviews and don’t directly write the complete code for me.

4. Finally, let me reconstruct it independently. Experience comes from: design choices, debugging, failure, and review, not from the "knowledge point list".

7. Establish a personal wrong question booklet. Let Codex maintain this structure every time: ## Wrong question book ### Error I thought:... Actually it is:... ### Why is it wrong? Underlying concept:... Triggering scenario:... ### Correct mental model When encountering similar problems in the future, first judge:... ### Review questions

1. ...

2. ...

3. ... This will turn "mistakes" into accumulated experience.

Open original on X