Applied Vibe Coding

IIT Madras · 15 Nov 2025, 3:30 pm IST · Remote
Anand S · LLM Psychologist · Straive
Transcript · Spreadsheet
CC0 - Public Domain

Set up GitHub

  1. Sign up at github.com
  2. Create a public repo, e.g. vibe-coding-workshop
  3. Click on the "LICENSE" button and add a license, e.g., MIT

Now you have a GitHub repo to store your code!

Fill up your profile in this spreadsheet.
It will look like https://github.com/your-username/

Use an online coding agent

You can also try with a CLI tool or VS Code extension

npx -y @github/copilot-cli       # GitHub Copilot CLI
npx -y @openai/codex             # OpenAI Codex
npx -y @anthropic/claude-code    # OpenAI Claude Code
npx -y @google/gemini-cli        # Google Gemini CLI
npx -y opencode-ai               # OpenCode AI

There are many others. Most of these have Visual Studio Code extensions too.

Publish any app. Here are some ideas

  1. Create a game, e.g., crosswords, memory, quizzes, puzzles
  2. Create a tool, e.g. synthetic data generator, Excel to JSON, password strength checker
  3. Build a web scraper, e.g., news, weather, stocks, sports, movies
  4. Analyze data, e.g., ratings recommender, logs analysis, sentiment analyzer, geospatial analysis, solve Kaggle contest
  5. Build a dashboard, e.g., fitness tracker, sales analysis, social analytics, project management, bug tracker

Publish any app. Here are some ideas (contd.)

  1. Generate content, e.g., social media post, invoice/receipt PDF, meme generator
  2. Do real-world tasks, e.g., reply to RFP, meeting TODOs, fill claims
  3. Build an API, e.g., URL shortener, auth service, webhook receiver
  4. Chrome extension, e.g., ad block, dark mode toggle, price track
  5. Mobile app, e.g., recipe manager, flashcard, share with clipboard
  6. Desktop app, e.g., note-taking, to-do list, file organizer

Some lessons to keep in mind

  • Don’t promise. Deliver, then share outcomes
  • Specs are the product. Code is a build artifact
  • Optimize for review, not coding
  • Restart on failure. It's faster than correction
  • Make repos LLM-friendly
  • Work in parallel
  • Expect uneven gains: seniors benefit differently than novices

Don’t promise. Deliver, then share outcomes

AI capability is uneven and unpredictable across tasks (“the jagged frontier”).

Don't commit to your manager/client. Build it. Then share it.
Fixing is 10X effort.

Or, commit to broad improvement, not specific apps.

Specs are the product. Code is a build artifact

AI makes typing cheap; the bottleneck becomes the specification.

Provide structured instructions.

Commit your prompts (e.g. SPEC.md, PROMPTS.md).

Optimize for review, not coding

AI coding agents tools boost development speed speed, shifting effort to verification.

Ask agents to write readable code.

Ask it to write test cases, run them, and ensure they pass.

Restart on failure. It's faster than correction

Once a conversation goes off-track, fixing all the mistakes is often more work than starting over.

Try once, maybe twice. If it doesn't work well, start again.

Break the task into smaller chunks - that can help.

Make repos LLM-friendly

Models perform better with consistent structure, fast tests, and machine-readable context (e.g., repo "manuals" for LLMs).

Standardize your AGENTS.md.

Provide examples for them to copy from.

Work in parallel

Give an agent a task, then work with another ageint while waiting.

These can be on different parts of the same project, or entirely different projects.

This means you need a queue of coding agent friendly tasks ready to go.

Expect uneven gains: seniors benefit differently than novices

Experienced devs get sizable productivity boosts.

But novices can develop an illusion of competence without robust verification.

Plan to learn along with the agents, not just delegate.

Applied Vibe Coding

IIT Madras · 15 Nov 2025, 3:30 pm IST · Remote
Anand S · LLM Psychologist · Straive
Spreadsheet
CC0 - Public Domain