Takeaways from Bangalore's first Claude Code Meetup

I went to Bangalore's first Claude Code Meetup and heard Kashyap M from the Claude Code team share some insider takeaways. 👇

This meetup was very developer-heavy. But as someone who uses Claude Code extensively, I know there's an entire community of users from non-engineering backgrounds who'd benefit from a session like this.

If there were a meetup focused on Claude Code for non-technical users and power users, would you show up? Drop a comment if that's something you'd want. I might just make it happen. (You can register your interest here)

Below is a distilled summary of every substantive insight from the audience Q&A.

1. Claude Code's Internal Development Velocity Is Extreme

90%+ of Claude Code is written by Claude Code itself. No part of the codebase is the same as it was 6 months ago ("ship of Theseus").
Each developer ships ~5 PRs/day; the team does 150–250 PRs/month.
~100 internal releases happen every single day. Every merged PR triggers an internal release within 30 minutes. These are batched into a single external release the next morning.
Bug-to-fix-on-your-machine turnaround: <24 hours.

2. Dogfooding Is the Primary Quality Assurance Method

The team's #1 bug-finding strategy is using Claude Code daily to build Claude Code.
Philosophy: "I have a pain point → fix it → if others don't care about your fix, make the pain point generalizable."
They have a formal test suite but emphasize that real-world self-use catches the most issues.
Every regression fix is accompanied by a regression test, creating a positive feedback loop over time.

3. Agentic Search > RAG for Code Navigation

Claude Code uses agentic search (find, read, grep tools) rather than RAG with vector embeddings.
RAG: "Higher floor, lower ceiling" — quick initial results but harder to fine-tune (retriever weighting, hosting databases, etc.).
Agentic search: "Lower floor, higher ceiling" — slower ramp-up but easily tunable via CLAUDE.md files. Performance improves progressively as you add more context to these files.
In Anthropic's and customers' experience, agentic search outperformed every other search approach once properly configured.

4. Verification Loops Are the #1 Unlock for Agent Productivity

Kashyap repeated this three times as the single most important insight for agentic developers:
- "The biggest bottleneck for production agents is the verification loop."
- Agents can generate mountains of code but without verification, you don't know if it works.
Practical advice:
- Write tests before giving Claude a task. Make outcomes binary (pass/fail).
- State clearly in your prompt: "This is how you verify your solution."
- Claude will self-correct 99/100 times when it sees failing tests.
- This also combats the sycophancy/confidence bias that was especially bad in 3.7 Sonnet (now largely fixed).

5. Hooks Are an Underrated Power Feature

Hooks let you run custom code at specific points in Claude's lifecycle — e.g., when a session starts, when Claude asks for permission, when Claude finishes speaking.
Stop hook pattern: After Claude finishes, a hook sends a prompt back saying "This was the original task — have you finished and verified it properly?" This forces a self-verification loop without manual intervention.
Combined with verification, this is what Kashyap called "the hack" for getting Claude to do 3× more reliable work.

6. Plan Mode Is the Secret to High-Output Development

Shift+Tab twice enters Plan Mode → Claude asks clarifying questions → produces a ~500-line plan (problem, solution, verification steps) → you hit Enter → it executes.
Plan Mode's first option after exploration is "Clear context and accept the plan" — it wipes the exploration context and keeps only the plan file, dramatically improving context management.
This is how the team ships 5+ PRs per dev per day. "It's not rocket science beyond that."

7. `CLAUDE.md` Is the Backbone of Context and Structure

There's no file/folder structure that Claude inherently favors — but you should describe yours in CLAUDE.md.
~10 lines in CLAUDE.md describing where files, assets, and tests live dramatically improves navigation.
For planning, a single plan file suffices. If your plan spans multiple files, you likely need separate sessions instead.
You don't need to memorize best practices — just prompt: "Use the Claude Code guide sub-agent, look at my past 10 sessions, and give me tips to improve my workflow."

8. Multi-Repo / Cross-Service Reasoning Is Already Possible

Use /add-dir <path> to load additional repositories into the same session.
Claude reads all directories simultaneously, including their respective CLAUDE.md files.
This can be persisted across sessions so every new session auto-loads all configured repos.

9. The Figma ↔ Claude Code MCP Integration Is a Game-Changer

Figma's MCP allows Claude to fetch designs and convert them to code, and now also the reverse — code → Figma designs.
Internally at Anthropic, the design-to-product cycle time has dropped dramatically thanks to this integration.
MCP (Model Context Protocol) acts as a universal data connector for Claude.

10. Desktop App Advantages Over Terminal

The desktop app bridges remote and local sessions in a single interface.
It offers native work-tree support, letting you work on multiple tasks simultaneously.
Better file system access without terminal management overhead.
Phone-to-terminal access is something they're working on — desktop app is currently the best bridge for that use case as well.

11. What Developers Actually Do Now (Role Shift)

"Imagine you are a manager with as many junior engineers as you want — except 'junior' is misleading because Claude is smarter than most software engineers I've interacted with, including myself."
Developers now focus on: prompting, roadmapping, steering, and verification.
Claude is not great at long-term (3-month) roadmap thinking — humans still own that.
The goal: "For every minute I'm not using Claude, I'm losing 50 minutes of developer productivity."

12. Claude Code's Product Roadmap Is Engineer-Driven, Not Top-Down

The team has gone from 0 to ~1.5 PMs. Most features originate from engineers at the ground level.
There is no "this is what you work on for the next 6 months" mandate.
Claude Code generated over $1B in revenue run rate in ~6 months.

13. The Future Points to Always-On, Remote Autonomous Execution

"I don't think the terminal UI will last much longer in the way it is now."
The vision is remote, independent execution — Claude working on tasks autonomously, not waiting for human prompts.
"If every single person's Claude isn't working while we are at this meetup, something is seriously wrong."
They already have claude.ai/code — a web-based sandbox with repo cloning on Anthropic's servers.

14. Security and Alignment Are Multi-Layered

Security is handled at the research level (model alignment), harness level (proper prompting), and classifier level (harmful request detection).
Re: Opus 4.6 sabotage risk report — misalignment is a known concern; there's no perfect answer yet but it's "extremely high focus."
Practical user tip: Have another Claude session review your PR for security vulnerabilities.

15. Handling Slow Browser Automation (MCP + Chrome)

If Claude is slow when automating browser tasks (e.g., booking Airbnb): diagnose whether the individual action is slow (→ use a faster model like Haiku) or whether Claude is exploring (→ create a skill that tells Claude exactly where UI elements are).
Skills reduce exploration time by giving Claude prior knowledge of the page layout.

16. Context Management Tips

Be aggressive about clearing context — use Plan Mode's clear-context feature.
If you're going in circles, drop the session and start fresh.
Don't try to do too much in a single session. Multiple focused sessions > one sprawling one.

Disclaimer: This summary was AI-generated from a transcript of the session. While care has been taken to ensure accuracy, it may contain unintended mistakes or misrepresentations of what was said.