GenAI is fast changing the modus operandi of software development. But to build good code from AI, you need to combine traditional programming with the new skill of coaxing code from code, otherwise known as prompt engineering.
In the latest episode of Leaders of Code, GitLab Field CTO Lee Faus and Stack Overflow CEO Prashanth Chandrasekar discuss why syntax is becoming less important, how to build better prompts, and why AI should be treated as a learning tool, not a crutch.
Understanding legacy code has always been challenging. Developers need to understand the context of why the code was written and what it does. But AI-assisted development shows us that syntax isn’t the hard part anymore.
In the podcast episode, Faus explains, “The syntax doesn’t matter anymore. And we’re figuring out that the real power of computer science is people’s critical thinking and problem-solving skills. When we sprinkle creativity in, we’re able to do magical things with the computers in front of us.”
Faus suggests that developers should learn languages like Rust and C++ while nurturing AI prompting skills for embedded development as tech evolves to support connected environments with edge computing.
AI development is accelerating with new models like deep research from OpenAI and DeepSeek. Chandrasekar says, “It doesn’t take a lot of imagination to see where this is heading. You have to be building for where the puck is going.” He believes, “Super intelligent agents won’t be fallible, but they can give 80 to 90% of what you need very effectively.”
AI offers far more than automation – it’s a learning and skill growth tool. Faus believes developers who excel at GenAI today are figuring out how to connect and assemble information that can amplify developer capabilities rather than replacing them. Leveraging GenAI in this way can help drive business strategy, refine product-market fit, and prioritize features that support sustainable growth.
Until now, the industry has primarily focused on AI’s automation and productivity benefits. However, Recent research signifies that AI is getting used to assist vital considering and complicated problem-solving. Wharton Professor Ethan Mollick’s research with Procter and Gamble confirmed AI can turn out to be a teammate, not only a productiveness instrument. Mollick believes organisations that focus solely on AI for productiveness will miss the chance to assume and act greater, and should face a backlash from employees involved about job losses.
Immediate engineering is an rising talent which professionals can develop by pooling information about what works. Immediate engineering is just like the launch of Google when folks wanted to grasp the artwork of looking to get to the fitting info.
Exact prompting with the fitting info sources produces higher outcomes. Faus describes builders’ frustrations when GitLab Duo would not ship the code they want. This usually outcomes from obscure prompts that lack enough context, the place some builders mistakenly strategy AI like a search engine and anticipate good solutions from easy queries.
Understanding immediate sorts improves outcomes: system prompts set AI conduct whereas person prompts request particular actions. Retrieval-augmented technology (RAG) permits builders to entry exterior information from sources like Stack Overflow, offering verified, real-time info relatively than relying solely on the information the AI has been explicitly skilled on.
Immediate engineering ought to work aspect by aspect with conventional programming. Prompting can construction AI interactions with roles, duties, and constraints for constant outputs. In contrast to programming’s formal syntax and precision, it depends on pure language. This makes it extra accessible to a much less skilled engineer, however with a agency disclaimer that simpler isn’t essentially higher. As AI instruments enhance, immediate engineering shall be invaluable and conventional coding will nonetheless be important for high-performance methods.
AI coding assistants have developed mild years since autocomplete. They now assist the total growth cycle, streamlining base-level coding. Chandrasekar notes that skilled builders can automate easier duties to deal with advanced, mentally intensive initiatives. Exams from GitHub Copilot present early adopters completed tasks 55% faster.
But immediate engineering is a problem for engineers as a result of it calls for a brand new set of abilities. Efficient prompts have to bear in mind language, context, and technique to shepherd AI fashions towards helpful outputs. And to temporary AI, it helps in case you have expertise of what beauty like, and ideally, briefing others to finish coding duties.
In contrast to conventional programming, immediate engineering introduces variability that may sluggish adoption. Groups should rethink workflows and study new abilities. Research from McKinsey reveals that AI adoption typically stalls when organizations underestimate the necessity for human oversight and adjustment in immediate design.
Past the technical, immediate engineering reshapes how groups work with AI. Engineers have to undertake a steady studying mindset as AI fashions evolve. Chandrasekar believes many AI implementations wrestle with change administration and the “complexity cliff” of contemplating a number of variables. This wants professional information to find out what’s wanted earlier than briefing an LLM. Efficient content material supply all through the developer workflow ensures “What they’re constructing is grounded in reality.” Though many advocate for mastery of prompting as a vital developer talent, different consultants consider one of the best ways is to ask the LLM to write its own prompt.
Chandrasekar believes AI is a boon, however junior builders must be cautious to not use instruments as a crutch. They have to perceive the underlying ideas to troubleshoot points.
Faus advises builders to not depend on AI for every little thing, like commit messages: “You must absolutely perceive what you are pushing.” AI additionally creates alternatives and challenges for enterprises in mapping developer skillsets, as GenAI’s base customary of code is now comparatively excessive. However organizations that exchange junior engineers with AI threat long-term ache by stymying the expertise funnel.
Engineer Charity Main says software program is an apprenticeship business. It takes time and expertise to mature into an experienced engineer, and you have to observe. Writing code is the commencement dance in the direction of composing methods.
With pure language programming (NLP), Faus thinks that builders can embrace new ways of working to be extra strategic and collaborate with groups concerned in product, advertising, or documentation. Non-technical colleagues can higher evaluate technical docs and rapidly prototype concepts.
Current programming languages like SQL have abstracted growth farther from machine coding. LLMs take this abstraction a step additional with prompts changing advanced code. English, or your language of alternative, is the brand new programming language. The low-code and no-code motion is opening up new prospects for prototyping. Chandrasekar says, “A few of our greatest folks [at Stack Overflow] are in a practical space however can code.” Area consultants with coding talents can now take ideas by way of to implementation, like a advertising crew piloting a brand new pricing tier. “We have all now obtained this machine-powered tremendous intelligence.”
Greg Benson, Professor of Laptop Science on the College of San Francisco and Chief Scientist at SnapLogic, says there may very well be a complete technology that may develop useful computer programs with none formal coaching. They could lack the foundational information and abilities {that a} grounded pc science schooling brings, however he thinks there is a path ahead by including people within the loop to engender belief.
AI-first growth is much from flawless. GitClear’s research reveals AI will counsel legitimate code however struggles with reuse and modification. Overreliance can weaken maintainability and take a look at protection, as defined by GitClear CEO Invoice Harding within the Stack Overflow Podcast. Chandrasekar notes enterprise warning round AI adoption resulting from reliability considerations. Stack Overflow’s Developer Survey suggests builders are smitten by utilizing AI instruments, but solely 43% really feel assured about AI accuracy, with 31% remaining sceptical. Simply half {of professional} builders trusted AI code accuracy.
This may very well be as a result of skilled builders see that it is easy to generate code, but it surely wants an overlay of experience to develop good code. Like textual content technology, AI code can look believable however not often works in the way in which you anticipate and wish it to. It is a quicker horse, not an electrical supercar.
So what’s the answer? Immediate with objective and increase AI with high quality information sources to mitigate dangers by way of correct attribution. Stack Overflow’s Enterprise Product choices had been constructed to energy folks and AI with high-quality, human-validated information.
- Stack Overflow for Teams: Essentially the most correct and reliable information retailer in your AI options.
- Knowledge Solutions: Our subscription-based API service that gives steady entry to Stack Overflow’s 15+ years of community-curated information.
To summarize the dialog: Builders who embrace AI as a studying and problem-solving instrument will thrive. Those that use it as a shortcut will fall behind. The way forward for coding is not about understanding the fitting syntax, it is about understanding tips on how to assume.
