Interview

How to Work with Claude Code, Antigravity, and Codex in 2026: The Fundamentals of Vibe Coding

These days, everyone is talking about Claude Code, Antigravity, and Codex. People are comparing their features and learning new skills for their AI agents. We asked our expert, Sergii Bataiev, to explain the basics of how to "vibe code" the right way. This way, you can avoid wasting your token limits and make sure your project actually works.

Artificial intelligence
Meet the interviewee
sergii bataiev
Sergii Bataiev
Architecture & Technology Director
  • Sergii has over 20 years of experience in transforming engineering excellence and building high-scale distributed systems for Fortune 100 clients.
  • Currently leads the Software Architecture Office at ELEKS and spearheads AI adoption across the delivery structure to integrate Generative AI tools and approaches into the software development lifecycle.
  • Successfully scaled technology operations from 40 to over 140 engineers while architecting complex ecosystems that balance innovation with operational stability.

Let's start with the basics. What are the main challenges developers face when they first start using AI coding assistants like Claude Code?

Sergii Bataiev: The main problem is that people often have unrealistic expectations. Many believe they can simply type "make me an app" or "write the code" and get a usable result. That usually doesn't happen. Instead, they end up wasting tokens, especially in Claude Code, where tokens run out very quickly.

If you are using the $20-per-month plan, you should not expect too much. More realistic results usually start in the $100-$200 range. This is why many people look at Codex from OpenAI or Antigravity from Google as more affordable options. However, the basic ideas are similar on all platforms.

 

What do you mean by "vibe coding" exactly, and how is it different from traditional development?

SB: In 2026, vibe coding is no longer just a meme, it is a recognised paradigm shift in software creation. Andrej Karpathy came up with the term in early 2025. It describes a workflow where the developer primarily communicates intent to an AI agent rather than writing lines of code/logic themselves.

 

So, how would you describe the right approach? Where should developers start?

SB: I would start with specification and planning first, then make small iterations, tests, and code review. In other words, don't ask for code right away; start with specifications.

First, describe your idea and ask the LLM to pose questions until you've worked through the requirements and edge cases. Eventually, you'll end up with requirements, key decisions, data models, and even a draft testing strategy. Collect all of this in a special file spec.md, and save it. This is your specification.

Without this foundation, AI agents often hallucinate and do the wrong thing. They need clear boundaries and expectations, just like human developers do.

 

What’s next? How do you translate specifications into actionable work?

SB: Feed spec.md or product requirements document into your LLM and ask it to generate a project plan - break down the implementation into logical, small tasks or stages. I use ChatGPT, thinking the latest model for this, and occasionally a fast model. You can adapt any reasoning-capable LLM for this purpose.

Having precise specifications and a plan means the AI won't make things up or add features on its own - it will follow the agreed-upon framework and constraints.

Additionally, break work into small steps (tickets) and tackle them incrementally. Making fundamental, sweeping requests to AI all at once leads to model failures and confusion. Work step by step instead.

Blog post
Explore AWS' answer to AI governance: a framework that enforces human approval at every stage
AI-DLC Explained: How AWS Labs Brings Structure to AI Coding
Application development
Artificial intelligence

What kind of context does an AI agent need to work effectively?

SB: The AI needs to see:

  • Which files to modify
  • Which parts of the code to reference
  • Version requirements and architectural restrictions
  • Project pitfalls
  • Team conventions

This is where connecting your GitHub repo in Codex and Claude Code is really convenient - you can pull context from the project. You also need project constraints (versions, architecture, rules about what's forbidden), potential pitfalls (where things broke before, what nuances exist), and preferred approaches (how things are already done in your team).

If you need current library documentation, you can also connect MCP servers or simply paste documentation snippets manually. There are even specialised MCP servers and services for automating context packaging.

 

Are all AI models equally good at all tasks?

SB: Absolutely not. While one model might excel at coding, others might be better at explanations, refactoring or analysing logs. You can also use one model to review another’s output to find mistakes, spot blind spots, or get a second opinion.

Knowing which tool to use and when to use it is one of the most overlooked parts of working effectively with AI.

 

When should developers introduce agentic tools into their workflow?

SB: Once they have specifications and a plan. Claude Code, Codex, and Antigravity are tools (AI agents, AI CLI) that are best applied after you have your spec.md and ticket plan. Then the agent can:

  • Read project files
  • Make changes iteratively
  • Run tests
  • Fix errors based on results

Without specifications and tickets, agents often hallucinate and do the wrong thing.

 

What about testing & commits in AI-assisted development?

SB: Make sure to test and check quality at every step. You might want to create a checklist or testing plan for each stage. If you use Claude Code, ask it to run a test suite after each task and fix any errors. It's also a good idea to review the code yourself or paste it into another AI tool to double-check.

Never assume AI-generated code is correct just because it runs.

Moreover, don’t forget about commits (or save points), those are the safety net that let you fix AI mistakes and understand changes. When you've modified files and all tests pass, you make a commit, and Git remembers: what changed, when, and why (with your comment). This way you can:

  • Roll back if something goes wrong
  • View change history
  • Share changes with others (via GitHub/GitLab)
  • Cleanly separate work into steps

For experiments with agents, it's convenient to create a separate branch (or worktree) so you don't mix stable code with experimental code.

 

How to avoid repeating the same instructions over and over?

SB: To put it simply, establish model rules. Create an instruction file for your project that covers style guidelines, linters, prohibitions/restrictions, team conventions, and testing/building commands. Different tools may call this file different things (e.g, CLAUDE.md, GEMINI.md).

 

What advice would you give to teams just starting their AI-assisted development journey?

SB: Start small. Pick a well-defined feature or module, create a proper specification, and work through it methodically.

Successful AI-assisted development isn't about throwing prompts at a model and hoping for the best. It's about systematic preparation, clear specifications, incremental progress, and quality control at every step. Follow these fundamentals, and you'll get real projects instead of burning through tokens on confused AI output.

Skip the section

FAQs

What is vibe coding?

The term vibe coding took off in 2025, thanks to Andrej Karpathy. With this AI-powered approach, developers use simple prompts to tell the AI what they want to build. The AI handles the code, debugging, and tweaks, so teams can focus on big ideas instead of getting stuck in the details. Instead of writing every line, you guide the overall direction and let the AI do the heavy lifting.

What is Claude Code?
What is Google’s Antigravity?
What is Codex from OpenAI?
Talk to experts
Skip the section
Contact Us
  • This field is for validation purposes and should be left unchanged.
  • We need your name to know how to address you
  • We need your phone number to reach you with response to your request
  • We need your country of business to know from what office to contact you
  • We need your company name to know your background and how we can use our experience to help you
  • Accepted file types: jpg, gif, png, pdf, doc, docx, xls, xlsx, ppt, pptx, Max. file size: 10 MB.
(jpg, gif, png, pdf, doc, docx, xls, xlsx, ppt, pptx, PNG)

We will add your info to our CRM for contacting you regarding your request. For more info please consult our privacy policy

What our customers say

The breadth of knowledge and understanding that ELEKS has within its walls allows us to leverage that expertise to make superior deliverables for our customers. When you work with ELEKS, you are working with the top 1% of the aptitude and engineering excellence of the whole country.

sam fleming
Sam Fleming
President, Fleming-AOD

Right from the start, we really liked ELEKS’ commitment and engagement. They came to us with their best people to try to understand our context, our business idea, and developed the first prototype with us. They were very professional and very customer oriented. I think, without ELEKS it probably would not have been possible to have such a successful product in such a short period of time.

Caroline Aumeran
Caroline Aumeran
Head of Product Development, appygas

ELEKS has been involved in the development of a number of our consumer-facing websites and mobile applications that allow our customers to easily track their shipments, get the information they need as well as stay in touch with us. We’ve appreciated the level of ELEKS’ expertise, responsiveness and attention to details.

samer-min
Samer Awajan
CTO, Aramex