Sunny Purewal

[Sunny Purewal @__lepton__]

@VictorTaelin Remove the first four words and you have a self improving loop towards an unknown goal. Magic.

[Sunny Purewal @__lepton__]

@gunnarmorling @grok what’s this about

[Sunny Purewal @__lepton__]

my CLAUDE.md file:

Do not make up unnecessary jargon.

that's it.

[Sunny Purewal @__lepton__]

@mosesxu @mattrickard I do the same PostToolUse hook for Linear MCP and have the session transcript attached to newly created Issues (which then get implemented separately). Transcript gets uploaded to S3 and linked in the ticket. Next hook only attaches the transcript since the previous one.

[Sunny Purewal @__lepton__]

everyone’s suddenly worried about AI Slop as if no one remembers Windows ME. Devs been sloppy since forever.

[Sunny Purewal @__lepton__]

@zeeg @henrytdowling in trying to reconcile “it needed to be said.” with “no one actually cares in the real world”

[Sunny Purewal @__lepton__]

The people who read code are probably the same ones that watch their agents code.

[Sunny Purewal @__lepton__]

so it turns out /compact is just deleting all the agent's messages from the history

[Sunny Purewal @__lepton__]

@astuyve agents need dev containers, no?

[Sunny Purewal @__lepton__]

@fchollet so often these are hidden deep in the implementation, how do we surface them to help understand the system? An architecture often "screams" its use cases not its edge cases.

[Sunny Purewal @__lepton__]

vibe coding and business problems don’t need to be in the same sentence. If you’re solving business problems you’re doing something closer to Harness Engineering. AI discourse has been taken over by business ambition and SaaS MRR when we could just be building software with extremely narrow use cases for a small user base, and that’s fine. We have tech companies building distributed systems to marginally help a pizza shop take orders and never stopping to think if anyone’s pizza eating experience is actually any better.

Custom software with extremely narrow use cases is the future. All the issues that come with scaling can be ignored if all software is bespoke.

[Kyle Gawley @kylegawley]

Vibe coding has revealed that people are completely clueless about solving real business problems

The software market is flooded with pointless apps that nobody uses

Shipping code faster is irrelevant

The bottleneck is and always will be talking to real humans to understand their problem, then build the right solution for a large enough group of users

That is the hard part, it has not changed and you can't automate it with AI

Stop bragging about shipping a useless app

Start bragging about how you spent 30 days talking to 100 users in-depth and uncovered a real problem that you are now solving for real money

[Sunny Purewal @__lepton__]

@ferrants love the analogy. This is why I’m hesitant to do “handoffs” between agents. Why have a separate review agent if the developer agent already built up the context to write the code?

[Sunny Purewal @__lepton__]

@dark_coderz because we've all accepted it's a risk.

"AI can make mistakes"

[Sunny Purewal @__lepton__]

@joshdholtz @mronge no hard requirement for code review in SOC 2

[Sunny Purewal @__lepton__]

@ben_sage delusion. if you think it's not already better than you.

Move past it, I don't even look at the code anymore, that's robot stuff now.

[Sunny Purewal @__lepton__]

@bendee983 There's no need to manage scalability if an org is generating a SaaS replacement for internal use that meets exactly the org's use cases. This is the crux of the SaaSpocalypse thesis, the LLM generating narrow use cases requires little maintenance, the org is their own customer.

[Sunny Purewal @__lepton__]

Take a deep breath and say it with me.

You don't need to review that code.

[Sunny Purewal @__lepton__]

@paulabartabajo_ > Running 10 parallel terminal sessions almost never makes sense IMO. The execution agent should run on a machine that is not the engineer's e.g. some central Background Agent. The engineer's LLM sessions produce issue tracker artifacts that are implemented by the agents.

[Sunny Purewal @__lepton__]

@aienginerd - Search for tool calls in the transcripts to link chats to e.g. Jira or Linear items

[Sunny Purewal @__lepton__]

@mitchellh @thatsFrScience does it involve multiplayer coding? 🤞