Addiction

paying attention. occasionally worth yours.

[Addiction @0xAddx]

Claude Code creator:

"Every time Claude makes a mistake, I don't tell it to do it differently. I tell it to write it to the claude.md or make a skill."

The same logic applies to prompts. The output is consistent because the input is consistent.

30 minutes to build. 10 hours saved every week.

Here's what the system looks like:

• content: posts, newsletters, repurposing. One template per format, one variable per task
• research: deep dives, competitor analysis, pre-decision research. Output that challenges, not reassures
• editing: full edit, ruthless cut, argument strengthening. Specific goals, specific constraints
• operations: emails, meeting prep, difficult conversation prep. Right every time
• decisions: decision framework, pre-mortem, getting unstuck. Recommendation at the end, no hedging

People open a new Claude session and start from scratch every time - while the people who get this fill in one variable and get output that needs almost no editing.

All 20 templates below

[Addiction @0xAddx]

https://t.co/Tt4tS7xHZe

[Addiction @0xAddx]

Boris Cherny the creator of Claude Code just shared the most important thing most people miss.

"The more context, the smarter the decisions will be."

He said it about code. It applies everywhere.

The same mistake kills advisory boards:

> advisors shaped entirely by what they've done before
> can't un-know what they know
> five people with five hammers looking for matching nails

5 Claude personas fix both problems:

> CFO builds three scenarios with real numbers
> CMO finds what the decision signals to the market
> Skeptic finds the assumption CFO and CMO both share and didn't question
> Customer tells you what buyers actually think
> Competitor tells you what rivals do in the next 30 days

Before: making the call alone at midnight with no one to pressure-test the thinking.
After: five perspectives, full context, synthesized into one recommendation, available in 30 seconds.

The article below is the full system

[Addiction @0xAddx]

@rileywestreel context windows matter more for reasoning than model size

[Addiction @0xAddx]

Boris Cherny the creator of Claude Code just shared the most important thing most people miss.

"The more context, the smarter the decisions will be."

He said it about code. It applies everywhere.

The same mistake kills advisory boards:

> advisors shaped entirely by what they've done before
> can't un-know what they know
> five people with five hammers looking for matching nails

5 Claude personas fix both problems:

> CFO builds three scenarios with real numbers
> CMO finds what the decision signals to the market
> Skeptic finds the assumption CFO and CMO both share and didn't question
> Customer tells you what buyers actually think
> Competitor tells you what rivals do in the next 30 days

Before: making the call alone at midnight with no one to pressure-test the thinking.
After: five perspectives, full context, synthesized into one recommendation, available in 30 seconds.

The article below is the full system

[Addiction @0xAddx]

https://t.co/fByDMy6xqB

[Addiction @0xAddx]

Boris Cherny, creator of Claude Code, said something that reframes how you should think about prompts.

In business, you have a principles document. Rules you lean on so you don't make ad hoc decisions every time.

The model can use those same principles. And they come in the form of skills.

Here's what he said founders should do right now:

> give everyone as many tokens as possible
> underfund projects slightly - 2 engineers instead of 4
> let them automate, they'll do it better next time
> cheaper next time, compounding from there

The individual version of this is a prompt library.

30 minutes to build. 10 hours saved every week. 5 categories. 20 templates. One variable per template. Everything else already written, already tested.

Most people open Claude and start from scratch every time. While the people who built this never write a prompt from zero again.

All 20 templates in the article below

[Addiction @0xAddx]

Boris Cherny, Claude Code creator, uninstalled his IDE after Opus 4.5. Not as an experiment. Permanently.

30 minutes to build. 10 hours saved every week.

Here's what the system looks like:

• content: posts, newsletters, repurposing, one variable per task
• research: deep dives, competitor analysis, pre-decision research
• editing: full edit, ruthless cut, argument strengthening
• operations: emails, meeting prep, difficult conversation prep
• decisions: decision framework, pre-mortem, getting unstuck

People open Claude and start from scratch every time - while the people who get this fill in one variable and get output that needs almost no editing

Read the full 20-template system below ↓

[Addiction @0xAddx]

@infobrA1n latency on synthetic feedback loops is the real competitive advantage

[Addiction @0xAddx]

Dario Amodei just asked the question nobody has a good answer to yet.

What are the moats when software becomes cheap and then the rest of knowledge work becomes cheap? We don't know. We've never quite asked that question.

He said there's going to be a huge scramble at the level of companies.

The same scramble is already happening at the level of decisions.

Most advisory boards look like this:

> three busy people who half-read the deck
> advice shaped by what they've done before
> equity charged for the privilege

5 Claude personas replace all of that:

> CFO builds three scenarios with real numbers
> CMO finds what the decision signals to the market
> Skeptic finds the assumption both CFO and CMO share and didn't question
> Customer tells you what buyers actually think
> Competitor tells you what rivals do in the next 30 days

Before: making the call alone at midnight with no one to pressure-test the thinking.
After: five perspectives, synthesized into one recommendation, available in 30 seconds.

The article below is the full system

[Addiction @0xAddx]

https://t.co/fByDMy6xqB

[Addiction @0xAddx]

Boris Cherny, Claude Code creator, uninstalled his IDE after Opus 4.5. Not as an experiment. Permanently.

30 minutes to build. 10 hours saved every week.

Here's what the system looks like:

• content: posts, newsletters, repurposing, one variable per task
• research: deep dives, competitor analysis, pre-decision research
• editing: full edit, ruthless cut, argument strengthening
• operations: emails, meeting prep, difficult conversation prep
• decisions: decision framework, pre-mortem, getting unstuck

People open Claude and start from scratch every time - while the people who get this fill in one variable and get output that needs almost no editing

Read the full 20-template system below ↓

[Addiction @0xAddx]

https://t.co/Tt4tS7xHZe

[Addiction @0xAddx]

@feraltekk context window constraints usually dictate how many agents collaborate

[Addiction @0xAddx]

@amaristw I like your approach

[Addiction @0xAddx]

Anthropic engineers compare prompting Claude to managing a new intern.

You have to be crisp. You have to be explicit. You can't assume it knows what good looks like in your context.

Jeremy from Anthropic's Applied AI team put it this way: if you were managing a new intern fresh out of college, how would you articulate exactly how to get things done? That's how you should think about giving instructions to Claude.

The problem is most people don't do this. They open Claude, type something approximate, get mediocre output, and start from scratch next session.

The prompt library fixes this permanently:

> one template per recurring task
> one variable per template, what changes each time
> everything else already written, already tested
> output is consistent because the input is consistent

30 minutes to build. 10 hours saved every week. 5 categories. 20 templates.

Most people spend more time writing prompts than getting value from the output.

The library inverts this permanently.

All 20 templates in the article below.

[Addiction @0xAddx]

https://t.co/Tt4tS7xHZe

[Addiction @0xAddx]

@amaristw how do you handle latency during prompt heavy workflows

[Addiction @0xAddx]

Anthropic engineers compare prompting Claude to managing a new intern.

You have to be crisp. You have to be explicit. You can't assume it knows what good looks like in your context.

Jeremy from Anthropic's Applied AI team put it this way: if you were managing a new intern fresh out of college, how would you articulate exactly how to get things done? That's how you should think about giving instructions to Claude.

The problem is most people don't do this. They open Claude, type something approximate, get mediocre output, and start from scratch next session.

The prompt library fixes this permanently:

> one template per recurring task
> one variable per template, what changes each time
> everything else already written, already tested
> output is consistent because the input is consistent

30 minutes to build. 10 hours saved every week. 5 categories. 20 templates.

Most people spend more time writing prompts than getting value from the output.

The library inverts this permanently.

All 20 templates in the article below.

[Addiction @0xAddx]

@0xRravix scaling systems for fifteen dollar rpm is smart leverage

[Addiction @0xAddx]

He works three jobs as a nurse in Norway.

He still found time to build a YouTube channel that made $31,785 in 90 days.

One to two hours a day to run it.

He posts a video every other day. Not because he has time. Because he outsourced the editing to a freelancer for $20.

Same price as two cheeseburgers.

AI writes the script. AI generates the voice. Freelancer assembles the video. He uploads.

No face. No camera. No studio.

$50-$200/month in tools. Finance niche pays $15-$40 per 1,000 views.

Month 3. $31,000/month.

The channel doesn't clock in.

The channel doesn't clock out.

[Addiction @0xAddx]

https://t.co/zuaowUxddW

[Addiction @0xAddx]

@0xRravix Okay
I won't be offended

[Addiction @0xAddx]

@rileywestreel cache retrieval latency is the real bottleneck for agent reasoning