AI Tools for Knowledge Work and Prompting

A field guide to getting real leverage from LLMs beyond code — prompt craft, deep research, NotebookLM, MCP servers, and generative image/video recipes.

The most useful thing about large language models is not that they write code — it is that they turn plain English into a general-purpose command line for research, analysis, and media. Andrej Karpathy's framing is that neural nets "have become programmable with large language models" and that this is a genuinely new kind of computer.¹ The user's own note on that talk lands it cleanly: we have moved "from writing strict recipes to letting machines learn from the examples to just talking to a helper that invents all steps by itself," and so "the hottest new programming language is English."¹ This page collects the highlights that treat that new language as a craft: how to prompt, how to point models at your own documents and tools, and how to coax images and video out of them.

English as the interface — and why prompting is a skill

If English is the programming language, then prompting is the syntax, and the highlights repeatedly show that the syntax matters enormously. McKay Wrigley's demo compares a default prompt to a specialized one and calls the difference "night and day" — the same model, drastically better output, purely from how you ask.² The pattern that keeps recurring is assign a role and a standard. Aditya Jha's learning trick is not "explain like I'm 5" but the opposite: "You're the senior dev. I'm the intern. Teach me how to build it like you'd expect it in production. Show me the real flow, gotchas, decisions, and best practices."³ God of Prompt's "Game Theory Strategist" does the same by casting the model as "a former Pentagon strategic analyst" before handing it a problem.⁴ Roles work because they compress a whole distribution of tone, depth, and vocabulary into one line.

Prompting is also recursive: the best way to write a prompt is often to ask the model to write it. Nghia Tran's early highlight — asking ChatGPT for "a better prompt for MidJourney to generate an iOS app icon with a bird" — is the whole meta-move in miniature.⁵ The deep-research playbook below extends it into an assembly line of prompts building prompts.

Anthropic's Opus 4.5 guide: prompting as tuning

The single most authoritative reference the user saved is Anthropic's own prompting guide for Claude Opus 4.5, "based on extensive internal testing."⁶ It reads as a set of dials rather than magic words:

The through-line is that prompting a frontier model is less about tricking it and more about removing the aggressive scaffolding earlier, weaker models needed — a maturation the user clearly noticed by flagging the whole thread.

The context principle: show the model everything

antirez's essay on coding with LLMs states the governing law of getting good output, and it generalizes far beyond code. His claim is that "the maximum quality of work is reached using the human+LLM equation" — not vibe coding, not full autonomy, but partnership.⁷ Two of his rules are really about context management: "Provide large context" and "Always show things to the most able model, the frontier LLM itself. Avoid any RAG that will show only part of the code / context to the LLM. This destroys LLMs performance."⁷ The insight the user highlighted is that hiding context to save tokens is usually a false economy; you want to be in deliberate control of exactly what the model can see.

Deep research: turning filings into theses

The clearest demonstration of leverage-beyond-code in the collection is @buccocapital's DoorDash walkthrough, where he uploads a company's S-1 and five years of 10-Ks and asks ChatGPT Deep Research to write an investment thesis.⁸ The method is a three-stage prompt cascade worth internalizing:

flowchart LR
    A["Ask the model to<br/>research best-practice<br/>deep-research prompting"] --> B["Turn that into a<br/>reusable prompt<br/>template"]
    B --> C["Ask the model, as an<br/>investment analyst, for its<br/>own analysis playbook"]
    C --> D["Merge template + playbook<br/>+ uploaded filings into one<br/>Deep Research prompt"]
    D --> E["3–5 page prompt →<br/>60–100 page<br/>cited report"]

He notes the template "routinely creates 3-5 page prompts that are generating 60-100 page, very thorough reports."⁸ The template itself is a checklist of prompt-engineering fundamentals — define the objective, gather context and boundaries, use specific language, give step-by-step guidance, specify output format, apply chain-of-thought ("think step by step"), set a role, request justification and references.⁸ Aditya Kondawar reports the same collapse in effort from the investing side: work that "used to spend 4-6 hours reading DRHPs during IPOs" now gets "80-90% of the job done" with a single Gemini 2.5 Pro prompt that starts by asking the model to "explain the business model to me like a 5-year-old."⁹

NotebookLM as a personal research studio

Where deep research reaches out to the web, NotebookLM turns your sources into interactive artifacts. Allie K. Miller's recap captures the shape of it: upload up to 50 sources — "notes, docs, videos, links, or pasted text" — and generate "summaries, study guides, FAQs, podcasts, or a Q&A bot in just a few minutes," free.¹⁰ Her use-case matrix spans students turning a semester of notes into a podcast, marketers turning "a 37-page lead magnet into a fun podcast," and researchers distilling a stack of academic papers into an FAQ.¹⁰ Simon pushes it further by combining NotebookLM Slide Decks with Deep Research to "create personal illustrated guides to anything" — his example is figuring out how to get his son started fishing in the Bay Area.¹¹ The pattern is grounding: the model reasons over a corpus you chose, not the whole internet.

MCP servers: giving the model your live data and tools

The Model Context Protocol is what lets an LLM stop guessing and start querying real systems. In the Cursor + Supabase workflow, MCP lets the editor "query Supabase directly to retrieve database schema in real-time, modify tables dynamically, and automate schema updates" — so "Cursor understands your backend natively without requiring extra input," configured with a small .cursor/mcp.json file.¹² The same primitive powers knowledge work outside coding. hrishikesh tried Zerodha's MCP server and got a "one-shot analysis of my entire portfolio vs nifty50 index for the past 4 years," concluding he could "literally build my own personal finance manager with few prompts."¹³ Chris Long finds the Google Analytics MCP "INCREDIBLY powerful for SEOs" — letting AI "gather analytics data, find top performing content, analyze your content strategy."¹⁴ MCP is the bridge that makes the "talk to a helper" model actually reach your accounts, databases, and dashboards.

Ingestion: getting the world into LLM-ready text

Before a model can help, the source has to be readable to it. Several highlights are really about this plumbing. Charly Wargnier flags Scrapeghost, a GPT-4 library that can "scrape structured data from HTML without writing page-specific code."¹⁵ Matt Pocock's open question — the "Best CLI for crawling a URL and turning it into LLM-ready text" — is the same need stated as a gap.¹⁶ Dan Shipper's early "how to turn any book into a chatbot using GPT-3" is the origin species of this whole genre: point a model at a corpus and query it conversationally.¹⁷ Karpathy's own note anticipates the endgame — that "going forward, things have to be accessible for LLMs," with docs turned into LLM-readable formats, because today's GUIs are all built for humans.¹

Generative recipes: images and video by prompt

The most recipe-like highlights are the generative-media ones, where a single reusable prompt reliably produces a look:

• Product ad grids — Amira Zairi's "one prompt to create an amazing Ad campaign": a "3×3 grid in 3:4 aspect ratio for a high-end commercial marketing campaign using the uploaded product as the central subject."¹⁸
• Figurines from selfies — Regy's Nano Banana Pro recipe: "Make a miniature, full-body, isometric, realistic figurine of this person, wearing ABC, doing XYZ, on a white background, minimal, 4K resolution."¹⁹
• Consistent restyling — Ege's JSON + Midjourney --sref setup to "retexture any image with the same style" of a shared style reference.²⁰
• Cinematic video — PJ Ace re-edited the Fellowship trailer "as if Studio Ghibli directed it," spending "$250 in Kling credits and 9 hours."²¹
• Interactive 3D with zero code — el.cine generated a gesture-controlled "3D product interactive app" with Google Gemini 3 + three.js and "zero coding."²²

The common structure is a template with slots: fix the style, format, and constraints; leave the subject variable. That is prompt engineering applied to pixels.

Applied leverage: from idea generation to hustle

The final cluster treats the LLM as an agent of applied action. Greg Isenberg's riff on Sam Altman's "era of the idea guy" is a 19-item playbook for mining ideas with AI — including "scroll your own chatgpt history… if you're repeating the same prompt weekly, that's a product hiding in plain sight," and searching Reddit for "is there a tool that" to find "users begging for tools that don't exist."²³ Gautam Shewakramani calls trip-planning "the real killer use case for ChatGPT in travel."²⁴ And Jackson Fall's viral experiment — giving "GPT-4 a budget of $100" and "acting as its human liaison" — is the earliest, rawest test of the model as a delegated operator.²⁵ The unifying claim, from Jack Raines, is why any of this matters: "When knowledge (or the ability to quickly acquire it) is table stakes, the only things that really matter are curiosity, creativity, and agency."²⁶ The tools commoditize the knowing; the leverage comes from what you point them at.

Related

• AI-Assisted Coding and Vibe Coding
• AI and Irreplaceable Human Work
• Productivity and Focus Systems
• Startups, Indie Hacking, and Business Strategy
• ChatGPT & OpenAI
• Claude Code (Anthropic)
• Cursor
• Overview