vaerx

AI & development

[vaerx @vaerxx4]

https://t.co/8IKHZjaqVA

[vaerx @vaerxx4]

MOST PEOPLE ARE ABOUT TO BURN THEIR FREE FABLE 5 WINDOW ON WORK SONNET 5 ALREADY HANDLES.

Most people:

-> ask Fable 5 easy questions all day
-> forget the July 7 cutoff
-> get quietly rerouted and never notice

People who read the setup guide:

-> save Fable 5 for the weeks-by-hand jobs
-> drop the routing rule straight into CLAUDE.md
-> plan the big migration before the window closes

Every detail is below.

[vaerx @vaerxx4]

MOST PEOPLE ARE ABOUT TO BURN THEIR FREE FABLE 5 WINDOW ON WORK SONNET 5 ALREADY HANDLES.

Most people:

-> ask Fable 5 easy questions all day
-> forget the July 7 cutoff
-> get quietly rerouted and never notice

People who read the setup guide:

-> save Fable 5 for the weeks-by-hand jobs
-> drop the routing rule straight into CLAUDE.md
-> plan the big migration before the window closes

Every detail is below.

[vaerx @vaerxx4]

https://t.co/8IKHZjaqVA

[vaerx @vaerxx4]

@cyrilXBT the schedule and sell part is doing all the heavy lifting

[vaerx @vaerxx4]

@cyrilXBT byok is the only way to compete with cursor now

[vaerx @vaerxx4]

@0xWast3 walled gardens aren't going to let you bypass them lol

[vaerx @vaerxx4]

@0xMortyx we are just babysitting the compilers at this point

[vaerx @vaerxx4]

@eng_khairallah1 wow, gift from Anthropic

[vaerx @vaerxx4]

@0xWast3 who's paying for power and maintenance though

[vaerx @vaerxx4]

@N01ennn the search interface shifts internally now

[vaerx @vaerxx4]

@thekuchh the 'set once' part implies a human loop still

[vaerx @vaerxx4]

@Bober_smart reach from a single video is insane

[vaerx @vaerxx4]

@ridark_eth it's always about who's liquid when things break

[vaerx @vaerxx4]

@kirillk_web3 context recall is the real hard part here

[vaerx @vaerxx4]

This is ANTHROPIC's own talk on getting the most out of Claude Code, from the engineer who built it.

Not theory. This is the exact process Anthropic runs on new hires: ask the codebase, don't touch it yet.

Ask the codebase → plan before editing → hand it your team's tools → give it a way to check its own work

Five habits, from simple to complex:

1/ Codebase Q&A first: it explores the repo itself, no indexing, no setup. Cut new-hire onboarding at Anthropic from 3 weeks to 3 days.

2/ Ask it to plan before it writes code: stops a 30,000-line request from turning into the wrong feature.

3/ Teach it your team's CLI and MCP tools once: it reads a --help flag and uses them from then on.

4/ Give it a way to check its own work, tests or screenshots: it iterates two or three times and gets close to perfect.

5/ Run several sessions in parallel over separate checkouts: the power users at Anthropic never run just one.

The key insight:

a feedback loop beats a bigger prompt. Give Claude Code a way to check itself, and the iterations do the rest.

Watch the full talk below.

[vaerx @vaerxx4]

@Atenov_D lol the atm test cuts deep

[vaerx @vaerxx4]

This is ANTHROPIC's own talk on getting the most out of Claude Code, from the engineer who built it.

Not theory. This is the exact process Anthropic runs on new hires: ask the codebase, don't touch it yet.

Ask the codebase → plan before editing → hand it your team's tools → give it a way to check its own work

Five habits, from simple to complex:

1/ Codebase Q&A first: it explores the repo itself, no indexing, no setup. Cut new-hire onboarding at Anthropic from 3 weeks to 3 days.

2/ Ask it to plan before it writes code: stops a 30,000-line request from turning into the wrong feature.

3/ Teach it your team's CLI and MCP tools once: it reads a --help flag and uses them from then on.

4/ Give it a way to check its own work, tests or screenshots: it iterates two or three times and gets close to perfect.

5/ Run several sessions in parallel over separate checkouts: the power users at Anthropic never run just one.

The key insight:

a feedback loop beats a bigger prompt. Give Claude Code a way to check itself, and the iterations do the rest.

Watch the full talk below.

[kaize @0x_kaize]

https://t.co/8PUvfLawTX

[vaerx @vaerxx4]

Anthropic CEO Dario Amodei:

"I met Ilya Sutskever. One of the first things he said to me was: 'The models, they just want to learn. You have to understand this.'

It was like a Zen koan. I listened, and I became enlightened."

The line became Anthropic's core training philosophy: clear the obstacles, give a model good data and room to operate, and it learns on its own.

Bookmark and watch ↓

[vaerx @vaerxx4]

THE TEAM BEHIND CLAUDE CODE JUST REVEALED IT SKIPS RAG AND INDEXING ENTIRELY.

Cal, Applied AI team, core contributor on Claude Code, explaining how the tool actually works under the hood.

No indexing. Claude just greps, globs, and finds its way through your codebase the same way a human developer would.

Three timestamps worth jumping to:

-> 06:50 why agentic search beats RAG entirely
-> 21:35 why power users run 4+ Claudes in parallel
-> 21:58 the escape-key trick nobody talks about

The 6-month roadmap to actually build agents like this is below.

[vaerx @vaerxx4]

@shmidtqq agencies are just packaging these tools now

[vaerx @vaerxx4]

@undefinedKi hard to argue with that

bookmarked

[vaerx @vaerxx4]

@rvaniaaaa the adversarial part is where the real value is