Latest / AI Fire Daily / #515 Neil: AI Skills Are The Secret Behind Loop Engineering In 2026
Transcript
- 0:00I don't prompt clod anymore. I have loops running
- 0:03that prompt clod. My job is to write loops. Beat.
- 0:07Man, a year ago that quote would have sounded
- 0:10like pure science fiction. Everyone was completely
- 0:12obsessed with prompt engineering back then. We
- 0:14were all just chasing the perfect sequence of
- 0:16magic words. Right. But the industry quickly
- 0:19realized that brilliant prompts fail without
- 0:22good context. So we naturally shifted our focus
- 0:25toward context engineering. We started handing
- 0:28the model the right files and history. Yeah,
- 0:30but handing over files is great for just a single
- 0:32task. And we are moving way past single tasks
- 0:35into full automation now. We are firmly entering
- 0:37the era of loop engineering. Exactly. So today
- 0:41we are taking a deep dive into what a loop actually
- 0:43is. We are also unpacking nine free AI skills
- 0:46you can install right now. You just grab the
- 0:48GitHub links and run them. Yep, you can install
- 0:50them today to build your own systems. But you
- 0:53cannot just tell an AI to work and walk away.
- 0:56You really need to understand the underlying
- 0:58raw materials first. For sure. So before building
- 1:01the house, let us define the bricks. A loop is
- 1:03essentially a small, highly autonomous worker.
- 1:07It actively finds specific tasks completely on
- 1:09its own. Right. And then it hands that work directly
- 1:12to an AI agent. It checks the final result for
- 1:15quality. Then it decides the next logical step
- 1:18entirely on its own. It does all of this without
- 1:20you typing a single thing. But those individual
- 1:23steps need strict, repeatable definitions. The
- 1:26AI needs clear constraints to function properly.
- 1:29That brings us to the core concept of an AI skill.
- 1:32We should clearly define that term for everyone
- 1:34listening. An AI skill is a reusable instruction
- 1:37file, making AI follow the exact same pattern.
- 1:40I like that. Think of an AI skill as a single,
- 1:43solid brick. The loop is the house you build
- 1:45from those bricks. Right now, most people are
- 1:47still trying to live in a single brick. Which
- 1:48is pretty limiting. We need to start stacking
- 1:50them together to build real systems. People also
- 1:53frequently confuse these individual skills with
- 1:55AI plugins. Oh, yeah. Huge difference there.
- 1:58A plugin is a massive, heavily bundled software
- 2:01package. It carries its own custom agents and
- 2:03dedicated servers. Whereas an AI skill is an
- 2:06incredibly lightweight text file. It just gives
- 2:08your model a predictable baseline behavior across
- 2:12sessions. Exactly. It just sets the ground rules.
- 2:14I get the appeal of not starting from scratch
- 2:16daily, but if an AI skill is basically a well
- 2:20-packaged prompt, Why go through the trouble
- 2:22of installing it? I mean, I could just paste
- 2:24my coding preferences into the chat box. Well,
- 2:27pasting rules manually degrades rapidly over
- 2:29long working sessions. The model slowly forgets
- 2:32your preferences as the context window fills
- 2:34up. Installing a skill bakes those instructions
- 2:37deeply into the system prompt. Oh, I see. Yeah,
- 2:40it guarantees the agent follows your constraints
- 2:42on token number 10 ,000. So it stops us from
- 2:45constantly re -explaining basic rules to the
- 2:47AI. Precisely. That system prompt injection is
- 2:50the foundation of loop engineering. Let's look
- 2:53at simulating a complete virtual engineering
- 2:55team. This brings us to a tool called GStack
- 2:58by Gary Tan. Gary Tan from Y Combinator. Right.
- 3:02GStack sets up distinct, highly autonomous engineering
- 3:04roles. You get a virtual CEO, an engineer, a
- 3:08reviewer, and QA. Yeah, and you summon the specific
- 3:11persona you need for the task. They actually
- 3:13communicate by passing structured JSON files
- 3:15back and forth. The CEO writes a product spec
- 3:18and passes it down. Then the engineer reads that
- 3:22JSON file and writes the code. Gary Tan claims
- 3:25he writes massive amounts of production code
- 3:27daily with this, up to 20 ,000 lines while running
- 3:30Y Combinator full -time. Which is wild! It is,
- 3:33but bulk generation usually just results in a
- 3:35bloated, unmaintainable mess. Does simulating
- 3:38a CEO persona actually change the quality of
- 3:40the raw code generated? That skepticism is totally
- 3:43fair when discussing bulk AI code. Standard AI
- 3:47agents just optimize for completing the immediate
- 3:49narrow coding task. The CEO persona fundamentally
- 3:52alters that underlying generation mechanism.
- 3:55Interesting. How so? It introduces a rigid constraint
- 3:58layer based on business value. It actively rejects
- 4:01code that doesn't serve the core product goal.
- 4:03Then the reviewer persona checks the engineer's
- 4:06work against the CEO's spec. So it forces the
- 4:08model to weigh the code against broader business
- 4:11logic. It stops the agent from building useless,
- 4:13overly complex features. Right, exactly. It provides
- 4:15top -down context, keeping the code focused on
- 4:18real business goals. Yeah, that top -down focus
- 4:21completely prevents endless sprawling feature
- 4:24creep. But, you know, generating new code is
- 4:27only half the battle here. Reading and understanding
- 4:29old code is often much harder. Oh, absolutely.
- 4:32I still wrestle with prompt drift myself when
- 4:34managing large code bases. Beats. Oh, we all
- 4:37face that mental drift eventually. Vyhov's Sysinti
- 4:41built an AI skill called Understand Anything
- 4:43to fix this. It completely reimagines the traditional
- 4:47developer onboarding process. Most AI tools just
- 4:50try to read the code line by line, but understand.
- 4:52anything flat out refuses to read code that way,
- 4:55it actually parses the abstract syntax trees
- 4:58of your project files. It maps out how every
- 5:01single module and function connects together.
- 5:03Then it generates a massive interactive visual
- 5:05dependency graph. You can physically see the
- 5:08data flowing through the application. That is
- 5:09huge. So it explains the code base exactly the
- 5:12way a senior engineer would. It breaks down the
- 5:14entire system architecture visually for a new
- 5:16hire. You grasp the massive interconnected picture
- 5:19in minutes instead of weeks. It is perfect when
- 5:23you inherit a messy project you never built.
- 5:25So your virtual team is coding efficiently and
- 5:28understanding the architecture. But efficiency
- 5:30does not always equal a great final product.
- 5:33Definitely not. Sometimes the AI output just
- 5:35feels completely lifeless and hollow. We all
- 5:38instantly recognize that dreaded robotic AI vibe.
- 5:42Yeah, those overly complex words and perfectly
- 5:44balanced but soulless paragraphs. We need specific
- 5:47skills to aggressively humanize the output. Hardik
- 5:50Pandya built a skill called StopSlop to cure
- 5:53this. StopSlop, great name. It strips out the
- 5:57repetitive patterns that make text sound artificial.
- 5:59Most people just tell Claude to sound more human
- 6:01in the prompt, but that generic instruction almost
- 6:04never actually works. Right, the model just ignores
- 6:06it. How does StopSlop mathematically force a
- 6:08different writing result? It actually targets
- 6:10the language model's actual token probability
- 6:13engine. Models mathematically favor words like
- 6:16delve or tapestry or robust. Stop -slop applies
- 6:19a massive negative weight to those specific tokens.
- 6:22So it mathematically forces the AI to choose
- 6:24the second most likely word. Yes. That simple
- 6:27constraint shatters the predictable robotic sentence
- 6:30structures instantly. It uses a strict set of
- 6:33rules to teach Claude about slop. Then it actively
- 6:36guides Claude to rewrite entirely around those
- 6:39tired patterns. It breaks the mathematical predictability,
- 6:42forcing the model to write naturally. Exactly.
- 6:44And that exact same mathematical predictability
- 6:47plagues user interface design, too. Anthropic
- 6:50engineers recently wrote a fascinating post about
- 6:53this exact phenomenon. AI models heavily default
- 6:56to template of looks when designing web pages.
- 6:59You almost always get the inter font on every
- 7:01single generation. You get purple gradients layered
- 7:04over a stark white background. The CSS animation
- 7:07is always kept to an absolute bare minimum. Always.
- 7:10Why do these powerful models mathematically default
- 7:13to safe? purple gradients. Well, they optimize
- 7:16for the absolute lowest possible design friction.
- 7:19The training data literally averages out millions
- 7:22of human web pages. The mathematical middle ground
- 7:24of the internet is a safe corporate gradient.
- 7:27They just average the internet and the web loves
- 7:28safe purple templates. Right. So to break this
- 7:31cycle, Anthropic released the front -end design
- 7:34skill. It forces Claude to commit to a specific
- 7:37style before coding. It separates the visual
- 7:40design logic from the structural HTML logic.
- 7:43It forces the AI to define CSS variables for
- 7:46colors first. There's also another brilliant
- 7:49tool called the taste skill. It gives you adjustable
- 7:51dials for the model's design parameters. Yeah,
- 7:54you can dial the design from calm and minimal
- 7:57to fully brutalist. It forces the AI to make
- 8:00bold, highly opinionated visual choices, it stops
- 8:04the model from constantly seeking the safest
- 8:06middle ground. We have to offer a very quick
- 8:08functional warning here though. A beautifully
- 8:10brutalist design that buries the actual product
- 8:12is useless. It still counts as a massive fundamental
- 8:16failure of design. Absolutely. Functionality
- 8:19always has to lead the design choices you make.
- 8:21Good design must always serve the underlying
- 8:23purpose of the page. So we have finally fixed
- 8:25our outputs and our interfaces. Let's talk about
- 8:28revolutionizing how we process our raw inputs.
- 8:31Research and organization are where these AI
- 8:32skills get really interesting. For sure. Managing
- 8:35large amounts of messy information is incredibly
- 8:37difficult. Graphify completely transforms how
- 8:40we handle scattered folders of notes. You just
- 8:42run a single simple command inside your directory.
- 8:46Andrej Karpathy posted a rough idea for this
- 8:48workflow on X. Just 48 hours later... Graphify
- 8:52appeared matching his exact description. How
- 8:55does it actually connect these completely random
- 8:57text files together? It uses a small language
- 8:59model to extract entity relationship pairs. It
- 9:02scans every single note looking for proper nouns
- 9:05and core concepts. Then it mathematically links
- 9:07separate files that share those exact entities.
- 9:10So it builds a beautiful connected visual knowledge
- 9:12graph from raw text. Notes become nodes and shared
- 9:15concepts become the connecting edges. It creates
- 9:18a spatial map instead of a flat linear list.
- 9:20Yeah. The launch post highlighted a staggering
- 9:23token reduction metric, too. You stop feeding
- 9:25the entire folder into the massive context window.
- 9:28You only feed the specific graph nodes you actually
- 9:30need right now. The numbers are absolutely staggering.
- 9:33Whoa. Imagine scaling to 71 .5 times fewer tokens
- 9:37per query. That is massive. Two secs silence.
- 9:40It completely changes the economics of managing
- 9:42a digital second brain. You can surface hidden
- 9:46patterns without spending a fortune on server
- 9:48compute. Which brings us to another incredible
- 9:51research tool for your workflow. Trying to keep
- 9:53up with industry trends requires endless tab
- 9:56switching. Last 30 days is a skill built by Matt
- 9:58Van Horn. It takes that entire messy research
- 10:01job completely off your plate. Yep. You just
- 10:04give the AI skill a single topic to track. It
- 10:07bypasses standard, sanitized Google search results
- 10:10entirely. It actively scrapes X, Reddit, and
- 10:14specific forums over a 30 -day window. Then it
- 10:16pulls everything into one clear, expert -level
- 10:19briefing document. But corporate websites usually
- 10:21offer highly polished, heavily sanitized data.
- 10:24Does pulling data straight from Reddit risk injecting
- 10:26too much noise into the research? Reddit definitely
- 10:29contains a lot of noisy, irrelevant garbage,
- 10:32but it also contains unvarnished troubleshooting
- 10:34and real human opinions. The skill uses a secondary
- 10:38AI agent just to filter that noise. It captures
- 10:41real -time human sentiment, balancing out sterile
- 10:43corporate web data. Exactly. It discards the
- 10:47meme responses and compiles actionable, verified
- 10:49user workflows. It finds the weird workarounds
- 10:53that are actually working right now, as one user
- 10:55perfectly described a specific AI skill online.
- 10:5930 days of deep research. 30 seconds of actual
- 11:03work. We're going to pause for a quick word.
- 11:05Yeah, this deep dive is brought to you by Datacamp.
- 11:08In a world rapidly moving from simple prompts
- 11:10to autonomous loops, Keeping your skills sharp
- 11:13is absolutely non -negotiable. Totally. Datacamp
- 11:16offers interactive hands -on courses covering
- 11:18AI, data science, and advanced machine learning.
- 11:21And you learn by actually doing. You write code
- 11:23directly in your browser. Start building your
- 11:24own foundational AI skills today by visiting
- 11:27datacamp .com. Alright, we have covered text,
- 11:30code architecture, and deep market research today.
- 11:32The final, most complex frontier for AI skills
- 11:35is video animation. Oh man, this is a massive
- 11:39technical showdown happening right now. We are
- 11:41looking at remotion going head -to -head against
- 11:43hyperframes. Both of these AI skills try to solve
- 11:46the exact same job. They turn a single text prompt
- 11:48into a real animated video. But they sit on completely
- 11:52different foundational computer science concepts.
- 11:55Remotion was launched back in 2021 with a huge
- 11:58community. Yeah, and it is built heavily on the
- 12:00React JavaScript framework. It controls every
- 12:03single video frame using complex React component
- 12:06code. It also relies heavily on extensive cloud
- 12:09rendering servers to work. HyperFrames takes
- 12:11a radically different stripped down path to video
- 12:14generation. The AI agent only writes plain HTML,
- 12:17CSS, and basic JavaScript. It avoids modern JavaScript
- 12:20frameworks entirely during the generation phase.
- 12:23I mean, I understand generating the basic layout
- 12:25code for a web page, but how does a plain text
- 12:27HTML file actually become an MP4? You cannot
- 12:31just play a text document in a standard video
- 12:33player. Well, the system runs a hidden, headless
- 12:35web browser in the background. It loads your
- 12:37HTML file and takes rapid -fire visual screenshots.
- 12:40It captures 60 high -resolution screenshots every
- 12:43single second. Wow! Then it uses a command line
- 12:46tool called ffmpeg to stitch them. It literally
- 12:49stitches thousands of screenshots together into
- 12:52a smooth video file. A viral post recently compared
- 12:55these two methods using Opus 4 .7. They fed the
- 12:59exact same animation prompt into both systems.
- 13:02People were left to judge the visual output completely
- 13:04for themselves. The HTML video looked dramatically
- 13:08smoother than the React video. Some developers
- 13:10actually wondered if HyperFrames was faking the
- 13:12results. They thought it was just a sneaky wrapper
- 13:15around their emotion engine. Bin Liu, the VP
- 13:17of Engineering at Hagen, answered that directly.
- 13:20He confirmed HyperFrames was built entirely from
- 13:22scratch using basic code. The core philosophy
- 13:25behind it is actually incredibly logical and
- 13:27elegant. AI models struggle to write complex,
- 13:29bug -free React video components. Right. But
- 13:31they understand basic HTML formatting and CSS
- 13:34animations flawlessly. React is ultimately just
- 13:37another flavor of JavaScript. Why is raw HTML
- 13:40so much easier for an advanced AI agent to parse
- 13:42than React? Mostly because HTML has structured
- 13:45the internet since the very beginning of the
- 13:47web. The model's training data contains billions
- 13:50of raw HTML pages. It has only seen a tiny fraction
- 13:54of that in specialized React code. HTML is foundational
- 13:58web data. The model has seen way more of it.
- 14:02Yeah, so letting the agent write what it knows
- 14:04best produces vastly superior results. You get
- 14:07much smoother text animations and cleaner vector
- 14:10logo reveals. Both skills work exceptionally
- 14:12well for short, punchy marketing videos. They
- 14:15save hours of tedious timelines scrubbing in
- 14:18traditional editing software. We need to pull
- 14:20all these separate concepts together now. We
- 14:22have covered a massive amount of technical ground
- 14:24today. We really have covered the entire modern
- 14:27development stack, from simulating a CEO to generating
- 14:29raw HTML video frames. Installing a single skill
- 14:32like GStack or StopSlop definitely saves time.
- 14:36But simply saving a few hours is clearly not
- 14:38the ultimate end game here. No. The real power
- 14:40unlocks when you start stacking these tools together.
- 14:42You link planning, QA, deep research, and front
- 14:45-end design into one flow. You transition from
- 14:47writing the actual code to managing the entire
- 14:49system. It is exactly like stacking Lego blocks
- 14:53of data. You call them automatically instead
- 14:55of explaining your needs from scratch. That interconnected
- 14:58stack forms the foundation a true autonomous
- 15:01loop runs on. Exactly. It handles the tedious
- 15:04prep work before you even open your laptop. Which
- 15:07leaves us with one final slightly provocative
- 15:09thought for you. Beat. It is the natural next
- 15:12step in this rapid evolution. If AI skills are
- 15:15the individual bricks and loops are the massive
- 15:18house, what happens to our daily work when the
- 15:20loops become smart enough to start coding and
- 15:23installing their own skills while we sleep? That
- 15:25question completely redefines the future landscape
- 15:27of human productivity. Thank you so much for
- 15:29taking this deep dive with us today. OTRO music.