Parth Amin

PM 9–5, founder 5–9. Building in public after work: + +

[Parth Amin @_aminparth]

What is Kloot?

[Kloot @klootgg]

Your OBS: six widgets, a chat bot, an extension.

Kloot: one browser source + a deck of buttons — giveaways, polls, lower thirds, hype — fired from your phone. Viewers join by scanning the screen.

60 seconds 👇

Free plan → https://t.co/GrmvMYojEH

[Parth Amin @_aminparth]

I'm building Kloot around this idea: one browser source in OBS, every moment on a button. You can test it on your own OBS canvas, no account:
https://t.co/0uIwoUvEQH

[Parth Amin @_aminparth]

Most stream overlays are built like dashboards. Everything on screen, all the time.

Viewers don't remember dashboards. They remember moments:
the death counter hitting 100
the giveaway wheel landing
the poll on your BRB screen

Fewer widgets. More buttons under your hand.

[Parth Amin @_aminparth]

@LeeWyattCorp By owning the audience instead of renting it.

Build in public, collect emails early, share raw build notes & failures, and solve one specific job a general AI can’t. A small group that knows you beats ranking for answers AI already gives.

[Parth Amin @_aminparth]

AI answers are eating the click. Here's how people who build after a 9-to-5 can still get found. https://t.co/G8KTwlWQiP

[Parth Amin @_aminparth]

@wakeona Exactly. Access is tracking trust, not org charts. The hard part is scaling that trust without scaling the risk.

[Parth Amin @_aminparth]

Goldman Sachs just narrowed who gets AI coding tools. That's not AI failing. That's a normal rollout step.

Per Bloomberg, updated internal guidance limits Claude Code and GitHub's coding tools to the bank's engineers. Most traders and bankers lost access. Strats keep it. The stated trigger: a review of productivity and risk concerns.

Most enterprise AI rollouts go through the same three stages:

1. Wide pilot. Give it to lots of people and see who actually uses it.

2. The real bill arrives. Not just licences. Code nobody reviews. Data that might cross lines it shouldn't.

3. Scope by role. Who can own what the tool produces?

That third question is the one that matters. An engineer who writes a script owns it: tests, review, deploy, on-call. A banker who generates a script that touches a model or a client deck usually has nobody reviewing it.

The answer isn't "no AI for non-engineers." It's giving them outcomes, not a terminal. Build the reviewed agent or internal app once, and let everyone use that.

Pulling a tool back from a group isn't failure. Rolling it out with no owner for the output is.

[Parth Amin @_aminparth]

Val Town is shutting down, and still: billing paused, migration guides written, code going open source.

How you close a product is the last feature your users remember.

[Parth Amin @_aminparth]

Amazon swapped a frontier model for a tiny decision model on one job and says it performed just as well at 98% less cost.

The job: label internal tickets into 20 categories.

That's the lesson for anyone shipping AI. A lot of your LLM calls aren't writing anything. They're picking from a list.

Two different jobs:

1. Generation. Open-ended output: summaries, drafts, answers. Pay for the big model.

2. Decision. A closed choice from a known set: route this ticket, tag this doc, pick the next step. You need a probability per option and a confidence score, not a paragraph.

Their numbers:
• ~$0.0222 per ticket on Sonnet 4.6 vs ~$0.0005 on the decider (~44x cheaper)
• ~4.2s per call down to ~0.9s

Caveats: it's their own blog, list prices only, and no accuracy table.

The part I'd steal is the boring one. They shipped a three-way switch: old model, both side by side, new model. Run both until the numbers agree, then flip.

Before you tune another prompt, audit your calls. Count how many are really multiple choice.

That count is your cost roadmap.

[Parth Amin @_aminparth]

Anthropic says one engineer gave Claude Fable 5.2 one task: design a prototype OS.

1,000+ subagents. 8 days. It boots on a laptop.

The build got cheap. The brief didn't.

One paragraph of intent now drives a week of work. Knowing what to ask for is the job.

[Choblin @choblin29]

This is INSANE. Anthropic says one of its engineers asked Claude Fable 5.2 to build an operating system from an empty folder.

The model coordinated over 1,000 subagents working around the clock for eight days and produced a working prototype OS that boots on a laptop, complete with disk, keyboard and network drivers, file storage, windows, a login screen, an installer and built-in Claude functionality.

[Parth Amin @_aminparth]

I put the full holiday calendar for indie shops (Nov 2026 → Oct 2027, with the post-by dates) here. Building ShelfSocial for exactly this:
https://t.co/hAxuiOLAVi

[Parth Amin @_aminparth]

One holiday, three Instagram posts. That's the rhythm for a small shop:

2 weeks out: the idea post (what to buy, why come in)
7 days out: the hours post
Day before: a Story reminder

US Thanksgiving is Nov 26.

So the idea post goes up this week, not the night before.

[Parth Amin @_aminparth]

1.5K people found Kloot on Google in the last 28 days 🤯

Streamers looking for an easier way to run giveaways, polls, and overlays are landing here.

One browser source. Fire everything from your phone. No chat bot.

Try it free → https://t.co/C2AZBRC7oQ

#streamingtips #obs #smallstreamer

[Parth Amin @_aminparth]

Before you open the AI chat, write the messy first draft yourself.

Not because AI writes badly. Because the draft is where you find out what you actually think.

Open the chat too early and you skip the thinking, then edit someone else's thinking.

Write first. Prompt second.

[Parth Amin @_aminparth]

The 4-step guide, with screenshots: https://t.co/l2WypIVotM

[Parth Amin @_aminparth]

I built Kloot so you never have to alt-tab out of a game to put up a BRB screen.

1. Add one browser source in OBS
2. Open the deck on your phone (web app, nothing to install)
3. Tap. It's on stream.

No pairing code. Your phone talks to the overlay, not to OBS.

[Parth Amin @_aminparth]

A PRD you can't explain with the AI closed isn't a PRD. It's a transcript.

[Parth Amin @_aminparth]

Anthropic had an agent build a wiki from 115 KB of docs. It came out at 203 KB.

If your summary is longer than the source, it isn't a summary. Most status decks have the same problem.

[Parth Amin @_aminparth]

From Anthropic's write-up on agent memory. Also worth stealing: "save what the screen can't show". https://t.co/LcMLZN42JU

[Parth Amin @_aminparth]

Anthropic tested how AI agents should use memory.

Strict rules like "only record failures" made the strongest models worse. What worked: tell the agent what memory is for, not how to use it.

Same lesson as managing good people. Give the goal, skip the procedure.

[Parth Amin @_aminparth]

Source, with the full tables (easier and harder test sets): https://t.co/eKZDs57upi

[Parth Amin @_aminparth]

The most useful skill in an AI system isn't being right. It's knowing when it might be wrong.

Hightouch published a test this week. A cheap model judged every record, and only the ones it was unsure about went to Claude Opus 5.5 for a second opinion.

On their easier test set:
Opus alone: 99.0% accuracy, $3.86 per 1,000
Cheap model + Opus when unsure: 98.8%, $0.88 per 1,000

About 22% of cases went to the big model. Roughly 4x cheaper for almost the same accuracy. On the harder set the saving fell to under 2x, but the hybrid still matched Opus.

The part most teams skip: this only works if the cheap model's confidence means something. One cheap model they tested was well calibrated, so most of its mistakes came with low confidence. The other was overconfident, and its hybrid fell behind on the hard cases.

What I'd take into any enterprise AI build:

1. Don't pick one model. Pick a routing rule.

2. Test calibration, not just accuracy. When it says 90% sure, is it right 9 times out of 10?

3. Treat the escalation band as a product decision. It's a cost vs quality dial you can explain in one slide.

4. Log every escalation. That's your list of hard cases, and your next eval set.

A cheap model that knows when to ask for help beats an expensive model you can't afford to run on everything.