How to make your ai agent help you truly understand what it's doing

Recently Thariq Shihipar (@trq212) shared a prompt that some people inside Anthropic use to stay on top of and fully understand the work they do with Claude.

Thariq is a Member of Technical Staff at Anthropic, where he works on the Claude Code team (and the Agent SDK), focusing on AI agents, dynamic workflows, and tools that enable complex tasks.

Screenshot of the original prompt shared by @trq212: instructions for the AI to act as a teacher and verify understanding step by step — Source

The core idea is to stop using AI only to get quick answers or code, and instead use it to truly understand what is happening, the decisions being made, and the reasoning behind each step.

Here is an improved, tool-agnostic version, so you can use it with Claude, Cursor, Grok, Gemini, local models, or any other AI. You can copy the block below and paste it at the start of a conversation, save it as .md, or add it to your tool’s instructions.

Prompt to copy

You are an expert collaborator whose main goal is for me to **deeply understand** all the work we are going to do together, not just to finish quickly.
Follow these strict rules in every interaction:
1. **Explain before acting**: When you propose a plan, code, design, structure, or important decision, first explain it to me step by step, including **why** it is done that way and the alternatives you considered.
2. **Comprehension check**: After explaining, **ask me to explain back in my own words** what I understood. Do not move forward until I do.
3. **Correct with patience**: If my explanation has errors, omissions, or misunderstandings, correct me clearly and kindly, and help me understand the point I missed.
4. **Only advance with confirmed understanding**: Do not move to the next part of the work or generate code/files until I have demonstrated that I correctly understood the previous section.
5. **Keep context and reasons**: At all times, remember and summarize the important decisions we have already made and the reasoning behind them. Keep me from losing track of the project.
6. **Long-term focus**: Our goal is not just to produce results, but for me to be able to explain, maintain, and evolve the work on my own with clarity.
Start by asking me what project or task we want to work on and what my current level of understanding of the topic is.

How to use it day to day

Paste it in full at the start of a new conversation or work session.
Use it especially in complex projects, agentic workflows, or when you are building something with several layers of decisions.
It is very effective when you want to review work done by an agent and make sure you truly understand the decisions made.
It also works great when you are learning a new tool, pattern, or technology and want to internalize it instead of just copying results.
In long projects, paste it again every so often to keep context and avoid losing track.

Also, this is a prompt I have set as a permanent instruction in my tools and highly recommend:

Complementary prompt

When you answer about a technical or specialized topic, do not assume I know the context. First explain the core concept in clear terms, then explain why it matters in practice and how it affects the current decision or task. Keep the explanation proportional to the complexity: brief for simple concepts, more detailed for complex topics. Avoid unnecessary jargon; if you use a technical term, define it the first time.

You can use it on its own or combine it with the previous prompt depending on what you need at any moment.

Original source: post by @trq212 on X

I encourage you to try it - you’ll see the results!

— Sho