Hardeep

AI [email protected] | sharing Claude workflows | IIT Delhi

[Hardeep @rav4nn]

Day 1: I had zero tech background. A friend told me to try Claude Code.
Day 60: ₹1L/month contract with a UK AI company.
Here's every project I built in between 🧵

[Hardeep @rav4nn]

Fix was a small background watcher that drops a .done file the moment the result file appears. The agent watches for that instead of guessing.

Stalls went to zero.

[Hardeep @rav4nn]

One of my coding agents hung for an hour last night. Not crashed. Hung.

It was waiting on a review step that never said it finished.

[Hardeep @rav4nn]

The setup: the agent writes its result to a file, a separate reviewer checks it. The agent was waiting for a "done" signal nothing was sending.

[Hardeep @rav4nn]

Building alerting across two environments taught me the real risk isn't a missed alert.

It's a prod alert landing in your dev channel, and nobody trusting the system after that.

So before any feature: a guard so prod events can't post to dev. 55 tests on that one rule.

[Hardeep @rav4nn]

My agents can write code now. Writing it was never the problem. Trusting it is.

So I built a gate: when Claude tries to commit its own code, a hook stops it and sends the diff to a different model first. No review, no commit.

One model writes. Another signs off.

[Hardeep @rav4nn]

@adxtyahq Use an SLM or a smaller model, faster LLM to convert the pre-final result into the validation layer.
This is the classic deterministic vs probabilistic issue with using LLMs, very few reliable ways to get through it.

[Hardeep @rav4nn]

A client wanted to know the moment their app broke, without drowning in alerts.

Backend through one tool, frontend through another. The rule that mattered: only alert on what's new or just came back.

Alert fatigue kills monitoring faster than missed errors do.

[Hardeep @rav4nn]

Why doesn't every podcaster have this tool already?

[Vipul Bhavsar @vipbhavs]

I started listening to Teen Taal last year, and it quickly became one of my favorite podcasts.

For those who don't know it, it's an absolute banger by @kuldeepmishra , @kkstau and @007asifkhan

But after hundreds of hours of listening, one question kept coming back:

"What all have they said about a specific person, place, book, food, or topic over the years?"

That curiosity turned into a side project.

Over the past few months, I indexed 262 episodes and 71,000+ moments from Teen Taal to create a discovery app for the show.

Now you can:

🔍 Search anything
A name, city, book, movie, or even a half-remembered Hindi phrase. Instead of just finding an episode, you'll get the exact moments where it was discussed, complete with transcript snippets and one-click playback.

💌 Browse TT Staff Letters
One of my favorite parts of the show. Explore letters by sender, see where they were read, and jump straight to those moments.

🎙️ Full episode pages
Transcript and player side-by-side. Click any line in the transcript and start playing from that exact point.

✂️ Shareable clips
Found a great moment? Share it directly.

📚 Curated episode lists
Perfect for discovering classic Teen Taal rabbit holes.
Honestly, this has completely changed the way I listen to the podcast.

Would love for fellow Teen Taal listeners to try it out and share feedback.

Link in the first reply 👇

[Hardeep @rav4nn]

Why are gyms always THIS empty on Sundays???

[Hardeep @rav4nn]

Market yourself more.
Make sure your work is visible to everyone.

It makes sure your higher ups can't take credit for your work.

[Arpit Bhayani @arpit_bhayani]

Be visibly competent.

If you have skills but others do not know about them, you will never get the opportunities you deserve. Make your abilities visible, and ensure that what you bring to the table is impossible to ignore.

[Arpit Bhayani @arpit_bhayani]

Be visibly competent.

If you have skills but others do not know about them, you will never get the opportunities you deserve. Make your abilities visible, and ensure that what you bring to the table is impossible to ignore.

[Hardeep @rav4nn]

@edenhazaa I paid and found out it's unwatchable on TV considering the low bitrate and 1080p stream. Even the highlights seemed noisy.

[Hardeep @rav4nn]

The pre-commit gate won't let anything through without a review receipt from Codex. Different model. Separate pane. Writes its findings before any commit lands.

One machine. One person. Seven projects moving at once.

[Hardeep @rav4nn]

Here's my actual terminal setup for running parallel Claude Code agents.

7 workspaces. 5 agents inside one project right now. A separate Codex reviewer as its own pane.

This is how I stay coherent across multiple projects at once.

[Hardeep @rav4nn]

Each agent gets a pane in the right column. They don't talk to each other. They write results to files. The orchestrator in the left pane reads them and decides what's next.

Context per agent stays clean. No single session accumulates 200k tokens of noise and starts making bad calls.

[Hardeep @rav4nn]

@naw103 Exactly!
After all the orchestration and decision making, context is too high for it to write code. Handing off to a separate agent to maintain a low context window becomes essential.

[Hardeep @rav4nn]

cmux + good guardrails is the answer to this problem
Keeping the context window clean will be the most important skill in 2026
It's easy once you set it up!

[aditya @adxtyahq]

Can someone please build this already?

An IDE that automatically switches models based on the task. Cheap models for simple edits, Claude/GPT for the stuff that actually needs reasoning.

And let me configure the routing rules myself

[Hardeep @rav4nn]

@adxtyahq Use a multi agent orchestrator like cmux and make the rules and skills yourself
Handoffs to different models based on use case is quite the need of the hour right now
Hit me up if you want to know how to set it up

[Hardeep @rav4nn]

@shrav_10 100% true. Youtube Music will pull bangers from your teenage years and play them just at the right moment.

Spotify has better playlists, though

[Hardeep @rav4nn]

if keyboard types all your code
what skill is actually yours?

[Tom ☕ @codevsdev]

if AI writes 80% of your code

what skill is actually yours?