“The real power of AI‑powered coding isn’t in replacing developers, but in amplifying our ability to solve problems across languages and domains.”
In recent years, I’ve leaned on large language models (LLMs) to automate small scripting tasks—everything from refactored Bash scripts to Python data pipelines. These quick fun achievements have kept me wondering: What if I could harness these models to accelerate real full‑stack development? Today I lead engineering teams through complex digital transformations. Yet my itch to build never went away. So, when an old MNIT Jaipur classmate asked for help on his startup’s SQL reporting pains, I saw an opportunity to scratch that itch—and explore how large language models (LLMs) could turbo‑charge full‑stack development.. What started as a simple SQL‑reporting utility blossomed into a deep dive into vibe coding, agentic workflows, and a renewed appreciation for human‑AI collaboration.
The Business Challenge
His operations lead needed ad‑hoc reports: log into a relational database, hand‑craft SQL, massage the output into CSVs. Simple—but manual, slow, and error‑prone. Could we embed a “smart” query assistant directly into his Vercel‑hosted dashboard, letting anyone ask natural‑language questions like:
“Show me total sales by region for Q1.”
…and instantly get a table, chart, or CSV?
Diagram of LLM‑powered pipeline as below
Picking the Stack—Pain Points Included
With years of back‑end muscle memory, I initially sketched a Python backend for the LLM logic, with a Next.js front end. But Vercel’s platform pushes you toward a single runtime. After wrestling with mixed‑language builds, I pivoted: all‑in‑JavaScript/TypeScript on Node.js.
The learning curve was steep. I had to:
-
Discover Vercel’s “v0 agentic mode” and its limitations (free‑tier quotas, usage warnings).
-
Get up to speed on shadcn/ui and Tailwind CSS for rapid UI prototyping.
-
Relearn Next.js conventions for server‑side API routes vs. edge functions.
By the end of Week 1, I had a skeletal “Table‑Stakes” project up on GitHub—and a burning question: How fast could I really go if I let an AI agent handle the plumbing?
Enter Vibe Coding
“Vibe coding” loosely describes a workflow where you direct an AI agent—via tools like Claude Sonnet 3.7 or ChatGPT—with short, intent‑based prompts, then iterate on its outputs in situ. It promised to:
Bootstrap boilerplate instantly
-
Generate utility functions on demand
-
Suggest best‑practice snippets (e.g., secure DB access)
…all without context‑switching between Stack Overflow, boilerplate repos, and your IDE.
But the ecosystem is messy: cline, Claude Code, Windsurf, Cursor—and each comes with its own CLI quirks, and not all work by default on Windows. No one had written a clear tutorial. So, I leaned on ChatGPT to draft my “starter kit” for vibe coding. I settled on Claude Sonnet 3.7 for agentic coding and VS Code for its rich extension ecosystem
Trials, Errors, and Quotas
A few lessons surfaced immediately:
-
Agent vs. API billing
Paid Claude Pro credits don’t apply to Sonnet’s agent API—unexpected costs ticked up quickly. -
Syntax habits
On Windows, early agent runs insisted on&&between commands. After a few painful debug loops, I explicitly prompted:“Use semicolons to chain shell commands on Windows.”
This worked mostly, but it still kept chaining them wrongly and fixing, wasting enough cycles as below:
-
Tool limitations
LLMs excel at fuzzy thinking (drafting logic, naming conventions) but can be weak in identifying the "right" thing to do. When faced with a challenge of compile time issue with a library that required older Node version, it kept trying to use--legacy-peer-depsinstead of upgrading the library version, even though when explicitly prompted it was able to see there is a newer version. Security and Testing as second grade citizens
The agentic mode created decent code but it completely missed adhering to security best practices or adding test cases and all of those had to be added separately. While it was able to add security controls, it had a really tough time adding unit tests. The following diagrams show lack of test coverage when code was generated and relatively severe security findings when I prompted and asked to fix:
Productivity Unleashed—With Caveats
With the environment squared away, I completed a working MVP in two nights—from DB‑query endpoint to a polished UI. Agents generated:
Next.js pages with form inputs
-
API wrappers for PostgreSQL
-
Client‑side chart components
All punctuated by manual fixes/ prompts for: testing, security, and cross‑platform quirks.
My take: AI agents can nudge you toward a “10× developer” pace—maybe even 100× in raw code generation. But they still forget to:
Sanitize user inputs against security vulnerabilities
Write comprehensive unit and integration tests
Handle edge‑case errors or rate limits gracefully
The following screenshot shows its attempts to fixing the issues when asked to.
Testing—Is It Cheating?
Generating test stubs from an agent feels like a shortcut:
“Write unit and integration tests for entire codebase using mocks as appropriate.”
Sure, it works. But can you trust tests generated against code that was AI‑assembled? My bias says: write only integration tests that validate end‑to‑end behavior, and manually review critical units. That way, you’re testing the system, not just the agent’s understanding of code it generated.
The Human in the Loop
Throughout, my role was less “coder” and more “orchestrator”:
-
Prompt architect
-
Context curator
-
Quality gatekeeper
I found that my domain expertise—knowing what the UX should feel like, understanding data‑schema trade‑offs, and recognizing security blind spots—was indispensable. The agent unlocked speed, but I guided purpose.
Where We Go from Here
AI‑assisted development is no longer science fiction. Its very much real, and very much in its infancy. Yet:
Non‑technical users still face too many moving parts to trust these tools in production.
-
Standardization (e.g., LSP‑style protocols for agents) is needed to bridge the “haves” and “have‑nots.”
-
Community knowledge (deep tutorials, case studies) lags behind hype‑cycle content.
Technical Domain knowledge will keep playing crucial role, e.g. difference in npm's dependency management vs maven's, or lack of Java like libraries like spring-data in nodejs will cause confusion until the ecosystems align
Fast evolving landscape: I started with a known debt of not using MCP. But in the two-three weeks since I started coding and now, significant changes have happened. Firebase Studio has been launched that makes it much easier, GPT-4.1 has been launched and is arguably better than Claude 3.7. Finally, Claude has published best practices for agentic coding.
Still, I’m pretty excited. If a backend‑focused engineer like me with increasingly higher management experience can become a lightweight frontend dev in a weekend, it shouldn't be very long until well‑tuned agent by our side could help us achieve very large productivity gains
Getting Started Yourself
-
Choose your agent: Claude Sonnet 3.7, ChatGPT Plugins, or GitHub Copilot.
-
Set up a single‑language stack on your hosting provider (e.g., Node.js on Vercel / v0, or Firebase).
-
Iterate with prompts—refine your instructions as you learn the agent’s quirks.
-
Guardrails first: add linters, input validators, and integration tests early.
-
Share your learnings: we need more deep‑dive tutorials, not just YouTube shorts or tweets.
Agents won’t replace us—but they will empower us to tackle more ambitious problems, faster. Embrace the rise of the AI‑assisted polyglot developer, and let’s build the future, one prompt at a time.
Check out the full project on GitHub:
https://github.com/AbhiramDwivedi/table-stakes
