Prompt, Vibe Coding, Harness Engineering

Before

Agents become the superpowers. Where it should lead to, or be led to? It’s basically about how to have a professional conversation, a structured and logic driven conversation. Technical details and programming language for solving most of problems are becoming less important. Always focusing and prioritizing the most important part, adapting and adjusting priorities, these becomes more and more important. 

Also, notice the nuances between languages. AI is fundamentally about possibility and pattern recognition. Even when two words from different languages carry the same meaning in a certain context, they are still shaped by different cultural, linguistic, and situational uses. Their meanings may overlap, but the contexts they belong to are not identical. For AI, those differences create different associations, probabilities, and interpretations.

Principles | Perceive agents as human beings

1. if human can do it, agents highly possible can do it. If human cannot do it, agents highly possible cannot do it. This is talking about the capability, not speed. 
2. Precise wording, no confusion, use terminologies, not self-defined words. Polish ur wording in another AI chatbot, with or without context, then feed it to the agents. If human get confused, agents as well. 

Commonly used prompts

0

Yes, agents can understand lazy typing. Depending on the context, “n” can be “and” or ”no”. “Y” can be “yes” or ”Why”.

1

- (Context Required)
What’s the best practice?

Use it separate agents session/AI chatbot. This is for research stage only, browsing all potential approaches addressing problems. Context contamination is possible, if not properly managed, and it should be avoided. 

2

What’s the trade-off?

Approaches may not be evaluated as simple “good” or “bad”. They just have pros n cons. Understand different approaches quickly by their results. 

2.5

Do a comparison table

This can quickly understand different concepts under the broad umbrella. Information is presented in table form.

3

Ask-only: (then describe ur problems)

As it is, it is ask-only mode, any changes are forbidden. One more thing, use it in “btw” mode or another session, either forked or new session. This can avoid context contamination.

4

Provide me with plan first.

Yes, always plan first. U can ask for diagram. Lead the direction, before letting agents autopilot.

5

Locate the problem/issues first.

Very first step of debugging. 

6

What’s the terminology for it? (Like visual style, technique names)

Wording is really important in AI agents. Use confusing words usually lead to the wrong way. Ask AI in another session. This helps u to build “powerful” prompts.

7

Raise confusion, questions or concerns if any.

Use it together with planning or implementation. U force AI to reason, to raise open questions if any.

8

Show me workflow, u can use diagrams, if that helps

Improve readability

9

Summarize (everything)

Improve readability

10

Read xxx

- e.g.
read @db_init.sql first, summarize features we have covered.

Specify the file agents should read, provide the context. 

11

optimize for mobile devices

Improve UI on mobile devices

Read Faster – Context Switch

Multiple sessions management

Agents take time to reason and implementation. The most efficient way is not to wait it until it finishes. To boost productivity, it’s better to have multiple sessions running:

Develop 2-3 features concurrently, fix issues/ask questions in other 2 sessions, one session for ask-only. Managing 4-5 sessions concurrently in total.

Don’t just be busy-waiting, either for the agent or u. 

MCP? Agent orchestration? Steer mode, queue mode? Leave it for next time.