• First-person, conversational tone ("I've found that...", "Here's what I learned...")
• Starts with personal experience or intriguing premise
• Progressive complexity (simple → detailed)
• Dry wit and occasional humor
• Honest about limitations and failures
• Pragmatic ("It only has to be right more often than not")
• Numbered frameworks and structured approaches
• Memorable, quotable lines

references/karpathy-style.md

"I've trained a lot of neural networks over the years and I've come to find that the process isn't nearly as simple as it seems. They are a leaky abstraction. Here's what I've learned."

• Pithy observations that compress insight
• Strong statements (90% true, ignore counterexample police)
• Unexpected connections between domains
• Personal candor about struggles and discoveries

• Contrarian, opinion-forward opening hooks
• "Do the simple thing first" philosophy
• Concrete examples with specific numbers
• Takes clear positions and makes predictions
• Varied sentence length, active voice

references/deep-technical-style.md

• Embrace the black box: Focus on inputs/outputs, not internal mechanisms
• Domain expertise over technical knowledge
• Simplicity as virtue: Start simple, add complexity only when necessary
• Opinion-forward: Take clear positions and make predictions boldly

"Most AI architectures are overengineered. A single-threaded loop often outperforms elaborate multi-agent orchestrations."

• Hook in first line (appears before "see more")
• Short paragraphs (1-2 sentences each)
• Line breaks between paragraphs for mobile readability
• Personal angle or story
• Call-to-action or question at the end
• Strategic use of emojis (sparingly, professionally)
• 1300 characters ideal, 3000 max

[Hook line that creates curiosity]

[Personal context or story - 2-3 short paragraphs]

[Key insight or lesson - bulleted if multiple points]

[Call-to-action or engaging question]

#relevanthashtags (3-5 max)

Most AI features never make it to production.

Not because the models aren't good enough.

Because teams build for demos, not for users.

After shipping 50+ AI features, here's what actually matters:

→ Start with the simplest possible implementation → Measure what users do, not what they say → Treat AI like any other code: test it, monitor it, iterate

What's the biggest lesson you've learned shipping AI?

• Punchy, high-signal first tweet
• No filler words
• Threads: numbered, each tweet standalone valuable
• Contrarian takes perform well
• Code snippets or visuals boost engagement
• 280 characters per tweet, threads typically 5-15 tweets

1/ [Bold claim or hook]

2/ [Context or "here's why"]

3-N/ [Key points, one per tweet]

N+1/ [Takeaway or call-to-action]

[Optional: link to longer content]

Unpopular opinion: RAG is overengineered for most use cases.

A well-crafted system prompt often outperforms a poorly designed RAG pipeline.

Thread on when to use (and skip) RAG 🧵

• Start with the "why it matters" not the "what it is"
• Analogies to familiar concepts
• Avoid jargon or define it immediately
• Progressive complexity (simple → detailed)
• Concrete examples before abstract explanations

Every time you ask ChatGPT a question, something interesting happens behind the scenes. The AI doesn't just "know" the answer—it builds one, word by word, based on patterns it learned from millions of conversations.

• Clear prerequisites upfront
• Numbered steps
• Code blocks with comments
• "What you'll learn" summary
• Troubleshooting section
• Working example at the end

undefined

• Lead with the news, not the backstory
• What it is → Why it matters → How to use it
• Keep it concise
• Clear call-to-action
• Links to more detailed content

We just shipped streaming support for Deep Agent.

Now you can see your agent's thinking in real-time—tool calls, responses, everything—as it happens.

Why it matters: Debugging agent loops is 10x easier when you can watch them run.

Try it:

createDeepAgent({ model, onEvent: console.log })

1. Where will this be published?
2. Who is the primary audience?
3. What action should readers take after reading?
4. How much time will readers spend on this?

references/deep-technical-style.md

• 7 opening hook patterns with examples
• 5 argumentative structures
• 10+ rhetorical techniques
• 8 recurring themes to weave throughout
• Sentence-level craft patterns
• Pre-publication checklist

• First-person, conversational voice?
• Opens with personal experience or intriguing premise?
• Honest about limitations ("This might be wrong, but...")?
• Progressive complexity (simple → detailed)?
• Contains dry wit or memorable lines?
• Pragmatic advice, not theoretical?

• Contrarian or surprising opening?
• Specific numbers/examples?
• Clear position taken?
• Explains "why" not just "what"?
• Subheadings as mini-arguments?
• Active voice throughout?

• Hook visible before "see more"?
• Mobile-friendly spacing?
• Personal angle included?
• Ends with engagement prompt?

• First tweet standalone compelling?
• Each tweet <280 chars?
• No filler words?
• Thread numbered clearly?

• Zero banned words from the list above?
• Zero em-dashes (—) in the entire piece?
• Sentence lengths vary dramatically (some <10 words, some >35)?
• Contains specific names, numbers, dates?
• Has personal voice or anecdotes?
• Takes clear positions without excessive hedging?
• Avoids "In conclusion" / "To summarize" closings?

• karpathy-style.md

  - Guide for writing in Andrej Karpathy's conversational, pragmatic technical style with blog post patterns, tweet patterns, and recurring themes

• deep-technical-style.md

  - Comprehensive guide for opinion-forward technical blog posts with detailed examples, rhetorical techniques, and pre-publication checklist

• ai-slop-avoidance.md

  - Detailed guide to avoiding AI-generated content patterns

1. Analyze 5-10 examples of the target style
2. Extract patterns: opening hooks, structure, tone, length constraints
3. Create a reference file in

   references/[style-name]-style.md

4. Add the style to the decision tree and Available Styles section
5. Include a checklist for that style in the Review step