Connective Blog

There is a buzzword in software called vibe coding

Andrej Karpathy popularized the term in 2025 to describe a style of AI assisted development where you do not worry about critically looking at the implementation. However I maintain that vibe coding is not new. It is what a developer does any time they work for more than 10 minutes without validating what they have created.

What is vibe coding really

When you are writing code guided more by this feels right rather than I have evidence this works you are relying on vibes. If you can sit and type for 30 to 60 minutes without compiling running pushing or testing anything you are not crafting you are improvising.

Validation can be anything that converts feelings into facts

• Running a unit or integration test

• Compiling and launching a quick manual check

• Observing logs while hitting a local endpoint

• Running a small script to sanity check performance or correctness

• Refactoring small pieces and then rerunning tests

• Pushing coding to a pipeline for automated validation

• Versioning working code

Without those feedback loops your inner mental model of the system is your only source of truth. You start stacking assumptions I am pretty sure this method returns X the DI is probably wired correctly that refactor should not affect anything important. You know vibes.

We already had a name for the alternative

Long before vibe coding the industry was talking about a more grounded deliberate approach to software. The Pragmatic Programmer by Andrew Hunt and David Thomas framed programming as a disciplined professional craft and warned against programming by coincidence while advocating ruthless testing feedback and adaptability.

The software craftsmanship movement picked that up and set out to put responsibility professionalism pragmatism and pride back into software development positioning itself as a continuation of the pragmatic programming ideas. Modern engineering playbooks from big vendors do the same tying maintainability to testability modularity documentation and clear boundaries.

What comes next is up to us, either we keep vibing or double down on being responsible in how we use AI.

Responsible Coding with AI

Responsible coding with AI means treating context as a first class concern. LLMs are only as good as the window you give them so part of the job now is writing compact purpose built documentation for the model tracking what has been done and making the state of the work explicit instead of just hoping the conversation remembers. It is the new Ctrl plus S era early computers could not be trusted to autosave and todays agents cannot be trusted to keep a clean mental model unless we feed them one.

Planning with your AI before coding goes a long way. The plan becomes the declaration of intention can be updated and serves as a where we were in case context gets corrupted.

Your agents also deserve the best tools. Creating servers and tools that use Model Context Protocol is how we hand LLMs power tools instead of asking them to whittle everything by hand. MCPs give LLMs access to repos databases CLIs and services in a structured and secure way. Not everyone has Claude Code or a fancy in house agent platform but everyone can ask on every task is there a tool or data source the agent is missing that can make this faster safer or more accurate. Building MCPs is easier than many think thanks to recent SDKs and of course LLMs themselves.

And underneath all of this automated tests still separate vibing from crafting. Test Driven Development is nothing new and others are arriving at the same conclusion. In an AI first workflow the code matters less than the test suite that defines what good enough means. If we are going to let models write more of the code we should be pushing them toward TDD style loops have the AI help write the tests then the implementation so we are always closing the gap between what we feel should work and what we can prove actually does.

About the Author

Cobbie Behrend - Software Developer with over 25 years of experience, most recently at Shopify helping teams increase their AI usage, and leveraging MCP's to automate reporting, auditing, and software development. Cobbie has a passion for using software to remove toil, and loves working with large painful legacy systems others might wish to avoid. He has completed multiple large migrations under budget and on time using custom tooling and test automation. Outside of work, he enjoys running long distances slowly on trails, and visiting his Daughter's family in Seattle.