I’m knee deep into AI coding agents and agentic programming. Yes, I have strong opinions about Vibe coding and I’m a big advocate of vetting your code. We are still responsible for our code, and AI doesn’t take accountability away from us.
Having said that, today I want to look at the bright side. I have really started enjoying my agents more. I found love in coding without typing code. This was not always the case - I always found auto-completion features more enjoyable than prompting in a terminal. But recently I have found my groove with prompts. So here are a few things I do now to feel like programming without actually writing much code.
Pseudocode with Keywords 🔗
I type in pseudocode using regular programming keywords - IF, ELSE, WHILE, SET, >, !=, ==, and so on. My prompts are mixed with TypeScript and Python with English.
IF transaction_id = null OR undefined THEN
ADD a rollback function
ADD a log entry
ADD error handling
CANCEL transaction
ELSE
FIRE Success Event and CALL consumer API
Constraints 🔗
I add a list of DOs and DON’Ts at the end of my instructions. I am able to express myself better and I am able to validate things fast, as my instructions are very targeted.
DO NOT - OVERRIDE EXISTING LOGIC
DO - CREATE NEW FUNCTIONS TO KEEP BEHAVIOUR MODULAR
Ask A Lot 🔗
I ask AI to explain the code, explain decisions, and what different approaches it would take. A lot of work goes into discussing code without actual implementation. Sometimes we throw away a big implementation once we find a better approach via this method by replacing it.
Explain the reasoning behind implementation
Propose another approach
Without editing any code provide alternative code change
What are major risks with your current implementation
Sub Agents and Context Management 🔗
AI responses are much better with context. Once the context window is filled, AI responses degrade. Sub agents are a great way to manage context outside your main task, keeping your main agent free to focus on the main task at hand.
Extensive Testing 🔗
AI can write good tests - AI can write bad tests. AI often writes silently passing tests. Read tests, run tests, change a few inputs in tests. Spend freed-up time in regression testing. We are moving from writing code to stewarding and steering it, and testing is the most important part of this new role.
Disclaimer 🔗
I use Codex and Claude mostly, but this should work with all coding agents. This is my process, and I do not guarantee it will work for all use cases. I do not write a full feature in one shot prompt - at least for now. I do not vibe code.
These minor rituals have helped me stay focused and produce high-quality code. I have started enjoying coding again, hopefully you will too. :)