Rahul

building software that builds software

[Rahul @chessarap1]

@AniC_dev One of the best products I’ve come across!

Using extensively for my AI tech stack

Keeping such low prices makes it affordable for economies like India. It’s gonna blow up a lot more!

[Rahul @chessarap1]

For the knowledge-sharing gap Scott mentions, put the alternatives you ruled out, and why, in the plan for teammates to review. Automated checks can pass without anyone else knowing why that design was chosen.

[Scott Berrevoets @ScottBerrevoets]

We now do a good chunk (60%+, potentially more) of automated code review with a bot to assess specific risks (arch, complexity, etc). Knowledge sharing is a problem but the rest is mostly working fine.

Influenced by what I wrote in March: https://t.co/Q2vxkXZxxp

[Rahul @chessarap1]

@JurajMalenica To follow the session itself, run /remote-control in Claude Code, then open it in the Claude app's Code tab. If you want Discord, Claude Code Channels has a Discord plugin for messaging a running session; it needs bot setup.

[Rahul @chessarap1]

There's a useful detail in Joseph's replies: his developer built the APIs, payments and AWS setup. He's building a mobile app on top. Getting excited to code again is valuable; it doesn't require doing all the engineering yourself.

[Joseph Carlson @joecarlsonshow]

Decided to jump into development again. Using Codex is how I imagine a kid feels playing minecraft. It's this incredible feeling that you can build literally anything you can think of.

This is the most fun I have had building stuff in a long time.

[Rahul @chessarap1]

@techenby For the pairing session, I'd bring a task you had to fix by hand. Talk through the plan and prompt together, then rerun from the same starting point and compare the diff with your fix. That gives you a concrete example to work through together.

[Rahul @chessarap1]

Kamshak's note suggests a useful test for shared agent tools. Have a teammate flag a problem in the artifact, then check whether that feedback makes it into the agent's next change.

[Kamshak @KamshakCH]

@AmpCode I recently shared a Claude Artifact with the team and commenting + sending stuff to the agent was a really nice experience. I'm not really sure what made it so nice but it felt very frontier. We're already using Amp multiplayer and I'm not sure why but it's less smooth

[Rahul @chessarap1]

@alliekmiller I'd connect slot 4 to the work in progress. For a coding agent, a 'needs help' ping could link to the running session so a teammate can inspect the current diff and help choose the next step.

[Rahul @chessarap1]

This deck has a useful split for agent-assisted work: agree on the plan and tasks as a pair, implement solo with the agent, then review the PR together. Pairing isn't just two people writing code.

[あまにい @amaniie_tw]

土曜日に登壇したXP祭り2026での資料アップしました！

ペアプロの価値はコードを書くことだけじゃない
https://t.co/DGeuT4FEfS
#xpjug

[Rahul @chessarap1]

In Nathan's July workflow, reattaching a terminal was for hands-on work with the agent; sharing a thread was for collaborating with teammates or clients. Worth separating those jobs when deciding what 'multiplayer' needs to do.

[Nathan Flurry 🔩 @NathanFlurry]

i've been loving amp specifically for bg agents because:

(a) if i need to iterate with it, i can re-attach a terminal to it in herdr/tmux for synchronous work as if developing locally

(b) i can share threads with other riveters or clients if collaborating on an issue, like vs code live share but for agents

[Rahul @chessarap1]

@SocksMyRocks @AmpCode A good first multiplayer test: take a real change, have one coworker call out an edge case, then trace it through the diff and test result together. That tests whether you can inspect and reason about the same change, not just whether everyone can open the same thread.

[Rahul @chessarap1]

@IamAroke Try pairing on a task where the design isn't obvious. Afterward, can both people explain the choice and what they ruled out? That's a useful check on whether the session helped both people understand the change.

[Rahul @chessarap1]

@Ojayy__x The middle ground is giving the second person a real job. Ask them to find a failing case or explain the design tradeoff while the first person drives. Then pairing can help with learning and decisions even when it isn’t faster typing.

[Rahul @chessarap1]

Yuki's team brought pairing back to read and accept AI-written code together. Faster code production hadn't removed the wait for human feedback. The talk treats lead time as the team's impression, not a measured result.

[yuki / ほにゃにゃ @honyanyas]

現地A 14:00-「AIとペアプロを始める。人とのペアプロをやめる。ペアプロの良さを改めて知る。もっと好きになった。」の登壇資料です！ありがとうございます！

ペアプロをやめたことで、ペアプロの良い部分を再認識し、もっと好きになりました！！
#xpjug
https://t.co/QnA6IoXLwz

[Rahul @chessarap1]

@GergelyOrosz If a change touches auth, permissions, or data handling, look at the agent's plan and commands before the PR is opened. Still review the finished diff; the earlier check doesn't replace it.

[Rahul @chessarap1]

These models on higher efforts like xhigh or max tend to over do things which are not required.
They need more instructions and strict steering.
Even Claude opus 5 on higher efforts does the same!

[Da7em @Da7_Tech]

I asked Sol to fix something ridiculously simple while I was away from my MacBook.

I came back to find it had burned through over 80% of my weekly limit because it opened a PR I never asked for.

Fable would never.

[Rahul @chessarap1]

Here is a link to it:

https://t.co/aq711knGnt

[Rahul @chessarap1]

I always ask Claude to review Codex’s work and vice versa

Thing is each model has its own strengths and particular things within the codebase

And I automate this asking part through a local AI agent harness I built so I don’t keep typing

[Daksh Gupta @dakshgup]

introducing model inversion

greptile detects if your PR was generated using GPT or Claude models, and uses the other model for review.

rohan and rodrigo from our research team share why we do this:

[Rahul @chessarap1]

@mattshumer_ How do you know it’s gpt 6??

There could be many more variants used internally in the 5.6 family?

[Rahul @chessarap1]

@shengkun_ye All your posts start with killing something

Why do you access data from something that you killed

[Rahul @chessarap1]

So approximately only 2 out of every 10 indie hackers ever make money that can put food on the table.

Sheesh.

[Marc Lou @marclou]

Last 30 days revenue distribution on @trust_mrr:

🪦 51% — $0 (pre-revenue)
☕ 35% — <$1K (coffee money)
🍜 10% — $1K–10K (ramen profitable)
🥩 3% — $10K–100K (steak profitable)
🛥️ 0.4% — $100K+ (yacht w/ AC)

n=8,281 startups