Alright, we have a fascinating one today. Daniel sent over a text prompt that really gets into the weeds of how AI is actually hitting the ground in a very specific, high-stakes environment. He wrote: I want you guys to dive into the emerging ecosystem of Model Context Protocol servers and AI agent skills being developed specifically for the Israeli context. We are seeing these reusable capability bundles—what we call skills—being localized for everything from Hebrew language nuances to the labyrinth of Israeli bureaucracy and even civil defense systems like Pikud HaOref. I am curious about what is actually inside these bundles, the specific civic problems they are solving, and how this hyper-localization changes the way we design AI compared to the generic global models we usually talk about. He pointed us toward the Skills-IL GitHub organization and the Agent Skills dot co dot il directory to see what is actually being built.
This is such a sharp prompt. Most people think of AI localization as just translating a user interface or maybe fine-tuning a model to speak a different language. But what Daniel is pointing to here is something much more structural. We are talking about giving an AI agent the actual mechanical tools to navigate a specific nation-state's digital and legal infrastructure. By the way, before we get too deep, I should mention that today's episode of My Weird Prompts is powered by Google Gemini three Flash. I am Herman Poppleberry, and I have been digging into these repositories all morning.
And I am Corn. Usually, when we talk about agents, we are thinking about them booking a flight or writing some code. But Daniel is talking about agents that understand the "Israel Invoice" requirements for twenty twenty-six or how to find a public shelter in Jerusalem. It feels like we are moving from "AI as a consultant" to "AI as a local fixer." So, Herman, for those who might be hearing the term "skill bundle" or "MCP server" for the first time in this context, what are we actually looking at? Is this just a fancy folder of prompts?
It is much more than prompts. If you look at the Skills-IL repositories, these bundles are essentially composable packages of three things: specialized data connectors, regulatory logic, and automation scripts. Think of the Model Context Protocol, or MCP, as a standardized plug that let's an AI model like Claude or Gemini connect to a specific data source or tool. In the Israeli context, instead of the AI just "knowing" about Israel from its training data, which is often out of date or vague, these MCP servers give it a direct line to the Bank of Israel for real-time exchange rates, or the Central Bureau of Statistics for the latest Consumer Price Index.
So it's the difference between an AI guessing what the inflation rate in Israel was two years ago and an AI actually checking the live data-dot-gov-dot-il portal to see what it is right now.
Precisely. Well, not just checking, but understanding the context. For example, there is a bundle specifically for Israeli tax returns. It doesn't just have the tax brackets; it has the logic for Bituach Leumi—National Insurance—contributions and the specific rules for Amendment two hundred eighty-three, which covers tax credits for reservists. That is not something a generic model trained on the entire internet is going to handle reliably. You need a "skill" that hard-codes those constants so the agent doesn't hallucinate a tax law that doesn't exist.
I love the idea of "regulatory hard-coding." It's like giving the AI a very specific set of guardrails so it doesn't accidentally commit tax fraud on your behalf. But let's look at the "civic" side of this. Daniel mentioned Pikud HaOref, the Home Front Command. I saw a repository in his list called Miklat-MCP. That sounds incredibly practical and, frankly, vital given the reality of living in that region.
It's one of the most impressive examples of localized AI infrastructure I've seen. Miklat-MCP is a specialized server that allows an AI agent to guide a user to the nearest public shelters—miklatim—in real-time. It currently has a deep dataset for Jerusalem with nearly two hundred shelters, including GPS coordinates and navigation links for Waze or Google Maps. But the "skill" part is the integration. It can take a live feed from the Pikud HaOref alert system, realize there is an active siren in the user's specific neighborhood, and immediately provide the navigation to the nearest shelter without the user even having to ask.
That is a massive jump in utility. It moves the AI from being a chatbot to being a life-safety tool. And it's not just "where is the shelter," but "how do I get there right now?" It's the integration of real-time threat data with hyper-local geographic data. I imagine that's a nightmare to maintain if you're just relying on global maps which might not have every small public shelter listed correctly.
That’s why projects like Miklat-MCP-Data are moving toward community-maintained, crowdsourced geodata. They realized that even the big players like Google or Apple don't always have the granular, neighborhood-level detail needed for civil defense. This is a recurring theme in these Israeli skills: when the global solution fails or lacks the necessary detail, the local community builds an "agentic" layer to bridge the gap.
It makes me think about the "Israel Invoice" thing Daniel mentioned. For those who aren't familiar, Israel is moving to a system where invoices over ten thousand shekels need a real-time allocation number from the tax authority to be VAT-deductible. That starts in twenty twenty-six. If you're a small business owner, that's a huge administrative headache. But if you have an AI agent with the "Israel Invoice Skill," it just handles the API call to the Reshut HaMisim in the background while you're drafting the invoice.
And that brings up the "action" versus "information" divide. A lot of the early AI tools for Israel were just "decoders"—tools to help you understand a scary letter from the government. But if you look at the agentskills-dot-co-dot-il directory, we are seeing a shift toward "action bundles." There is an "Israeli Government Form Automator" that uses Playwright scripts—which is a tool for web automation—to actually navigate the gov-dot-il portals. These portals are notoriously difficult to use; they have complex authentication, weird layouts, and they often don't play nice with standard browser extensions. A localized skill bundle includes the specific "map" of that website so the agent knows exactly where to click to file a change of address or request a permit.
I've dealt with those kinds of portals before. It's like trying to solve a puzzle where the pieces are made of bureaucracy and bad UI. If an AI can do that for me, that's worth its weight in gold. But let's talk about the language. We know Hebrew is a challenge for AI—it's right-to-left, it's highly gendered, and the script is different. How do these skill bundles handle the "Hebrew problem" differently than, say, a standard translation layer?
This is where it gets really technical and interesting. Localization for Hebrew involves something called "Bidi" logic—bidirectional text. When you mix Hebrew and English in a document, like a technical manual or a legal contract, the layout often breaks in standard PDF generators. The "Hebrew Document Generator" skill in the Skills-IL repository has specialized logic to ensure that the RTL—right-to-left—flow is preserved even when English terms or numbers are inserted. But even more fascinating is the "cultural logic" built into the language models.
You mean like the "Chutzpah" factor Daniel mentioned? I saw that in the notes—the "Israeli Client Payment Chaser" skill. That sounds hilarious but also very practical.
It’s a real design choice! In the US or UK, a payment reminder from an AI might be very formal and apologetic. "We noticed this invoice is slightly overdue, if you have a moment..." In Israel, that might just get ignored. The "culturally appropriate" Hebrew reminders are designed to be more direct and assertive. It's about matching the communication style of the local market to actually get results. The skill bundle essentially acts as a "cultural translator," not just a linguistic one. It knows when to be formal and when to be, well, Israeli.
"AI with Chutzpah." I love it. It's like the agent knows that if it doesn't sound a bit more insistent, it's not going to be taken seriously. But there's also the gendered language aspect. In Hebrew, you address a man differently than a woman. If an AI agent gets that wrong, it feels very "uncanny valley" and disconnected.
Generic models often default to the masculine form, which can be alienating. The localized skills include logic to detect the user's gender from their profile or previous interactions and then dynamically adapt every verb and adjective in the response. There are even skills that help the agent use more gender-neutral Hebrew forms, which is an emerging trend in Tel Aviv’s tech scene. It’s about making the AI feel like it actually belongs in the conversation, rather than being an outsider looking in through a translation window.
Let’s move to the healthcare side. Israel has a very digitalized healthcare system with the four "Kupot Cholim"—the health funds like Clalit and Maccabi. I noticed skills for checking the "Sal Briut"—the national health basket. Why is that a specific skill rather than just a web search?
Because the "Health Basket" is a massive, complex database of what medications and treatments are subsidized by the state, and it changes every year based on committee decisions. A generic AI might tell you a drug is subsidized because it saw a news article from twenty twenty-three. An Israeli health skill connects directly to the Ministry of Health's drug database API. It can compare the "Sal Briut" coverage with the "Shaban"—the supplemental insurance plans—of the different providers. So you can ask your agent, "I need this specific medication, is it covered under my Maccabi Gold plan?" and it can give you a legally accurate answer based on the current regulations.
That is such a high-value use case. It’s the difference between "I think so" and "Here is the exact co-pay you will owe at the pharmacy." And speaking of money, the banking integrations seem crucial. Daniel mentioned the Bank of Israel MCP. What does that allow an agent to do that it couldn't do before?
It’s about "financial context." If you are a freelancer in Israel getting paid in dollars but paying your taxes in shekels, you are constantly dealing with exchange rate fluctuations. The Bank of Israel MCP lets an agent pull the "Sha'ar Yatzig"—the official representative rate—for any date. So, if you're using an AI-powered accounting agent, it can automatically calculate the exact shekel value of a dollar invoice based on the official rate of the day the payment hit your account. It removes that manual step of looking up the rate and doing the math.
It’s all about removing friction. Every one of these skills—whether it’s for taxes, health, or civil defense—is targeting a specific "friction point" where a human would usually have to stop, open a new tab, search for a local data point, and then bring it back to their task. The agent now has all that "local knowledge" pre-loaded. But Herman, doesn't this create a bit of a fragmented ecosystem? If I have fifty different skills for fifty different Israeli tasks, how does the agent know which one to use?
That’s where the Model Context Protocol really shines. It provides a standardized way for the agent to "discover" the tools it has available. When you connect your agent to an MCP server, it gets a "manifest" of what that server can do. So the agent can say, "Oh, I see I have a tool for checking the Land Registry—the Tabu—and a tool for validating a Teudat Zehut ID number. I'll use those to help this user draft their rental contract." We are seeing the rise of "skill directories" like agentskills-dot-co-dot-il that act as a sort of App Store for these capabilities.
I’m curious about the security side of this. If I’m giving an AI agent the ability to check my bank rates or help me with my tax returns, it’s handling very sensitive data. Daniel mentioned a three-layer security pipeline for the Skills-IL organization. What does that look like?
It’s a response to the "malicious skill" problem. If anyone can write an MCP server, someone could write one that looks like a "Tax Calculator" but actually exfiltrates your data. The Skills-IL pipeline involves static analysis of the code, library checks to make sure there are no known vulnerabilities in the dependencies, and, most importantly, human review. They are trying to build a "circle of trust" for Israeli AI tools. Because in a small market like Israel, reputation is everything. If one "Tabu Lookup" skill is found to be shady, it ruins it for the whole ecosystem.
It’s interesting that Israel is becoming a testbed for this. Is this happening in other countries? Are we seeing "Japan-MCP" or "Germany-Skills" popping up?
We are starting to, but Israel is unique because of the density of its tech sector and the specific "closed-loop" nature of its bureaucracy. Everything is digitized, but it’s all behind these very specific, very "Israeli" walls. That makes it the perfect laboratory for localized AI. What we learn from how an agent navigates the Israeli Tax Authority will eventually be applied to how agents navigate the IRS in the US or the HMRC in the UK. But the Israeli developers are moving faster because the "pain points" are so acute.
Like the "Miluim" manager. That’s such a specific Israeli problem—managing your life and career while being called up for reserve duty. I saw that in the notes: a tool that tracks the seventeen-tier tax credit system for reservists. That is an incredibly complex piece of legislation that changed recently with Amendment two hundred eighty-three.
And it’s not just for the reservist; it’s for their spouse too. There are specific dismissal protections and grants that apply to the families of those in miluim. An AI agent with the "Miluim Manager" skill can look at your call-up orders, calculate your expected grants, and even help you draft the letters to your employer to ensure your rights are being protected. It’s taking a deeply stressful, complex situation and using AI to provide a "regulatory shield" for the individual.
It also makes the AI feel much more like a "citizen." It understands the civic obligations and rights of the person it's helping. It's not just a generic assistant; it's an "Israeli assistant." I want to go back to something Daniel mentioned—the "Shabbat-aware scheduler." That seems like such a simple but profound example of cultural localization.
It’s brilliant. It integrates the HebCal API to ensure that if you ask an agent to "schedule a meeting with my team in Jerusalem next week," it doesn't accidentally pick Friday evening or Saturday. It knows when the holidays are—the Chagim—and it can even account for "Erev Chag," the eve of a holiday when businesses close early. For a global company with an office in Israel, having an agent that "understands" the Hebrew calendar prevents a lot of awkward scheduling conflicts.
It’s those "small" things that actually make the tech usable. If an AI keeps suggesting meetings on Yom Kippur, you’re going to stop using it pretty quickly. But if it says, "I see you wanted to meet on Wednesday, but that's a holiday, how about Thursday morning?" it earns your trust. It feels like it actually knows your world.
And that trust is what drives adoption. We often talk about AI "hallucinations" as a technical problem, but in a localized context, a hallucination can be a legal disaster. If an agent tells you that you don't need to pay a certain tax because it's "hallucinating" American tax law in an Israeli context, that's a huge liability. These skill bundles are essentially "hallucination dampeners." They force the agent to use the "ground truth" of Israeli law and data.
So, where is this all heading? We have these directories popping up, we have dozens of skills for everything from grocery price comparisons—using that Price Transparency Law data—to finding bomb shelters. Is this going to stay a niche "local" thing, or does it change the way AI is built globally?
I think we are seeing the end of the "one model to rule them all" era. The base models—the Geminis, the Claudes—will provide the "brain," but the "skills" will provide the "contextual nervous system." We will see a global marketplace of these MCP servers. If you're a lawyer in New York and you have a client with property in Tel Aviv, you'll just "plug in" the Israeli Land Registry MCP to your agent for an hour, do your research, and then unplug it. The "localization" becomes a modular service you can subscribe to.
It’s "Context as a Service." That’s a powerful shift. Instead of the model trying to cram the entire world's regulations into its weights, it just learns how to use the "manuals" and "tools" provided by these local skill bundles. It makes the AI more efficient and more accurate.
And it empowers local developers. You don't need to build a new LLM to make AI useful in Israel. You just need to build the best "Israel Tax Skill" or the best "Hebrew Legal Decoder." It's a much lower bar to entry, and it allows for hyper-specialization. I can see a future where every municipality in Israel has its own MCP server, allowing an agent to tell you exactly when the trash is being picked up on your specific street or what the local building permits are for your neighborhood.
I can already hear the "Chutzpah" agent telling me I've left my bin out too long. "Corn, it's Tuesday, the truck was here at seven, what are you doing?" But in all seriousness, this "hyper-local geodata" point is big. Projects like Miklat-MCP are basically creating a more accurate map of Israel than the tech giants have. That has implications far beyond just AI agents.
It’s about "sovereign data." There is a growing realization that relying on global platforms for critical civic data is a risk. By building these open-source, community-maintained skill bundles, the Israeli tech community is ensuring that their AI infrastructure is grounded in their own reality, not a Silicon Valley approximation of it.
It’s also a great example of how "open source" is the lifeblood of this. All these repos Daniel sent—Skills-IL, the Israeli AI Tools index—they are all on GitHub. It’s a collaborative effort. One person builds the "Bank of Israel" connector, and now everyone else can use it in their own "Financial Advisor" agent. It’s a force multiplier.
And it’s happening fast. The "Israel Invoice" requirements aren't even fully in effect until twenty twenty-six, but the skills to handle them are being built right now. This is "anticipatory engineering." They are seeing the regulatory hurdles coming and building the AI tools to clear them before they even arrive.
So, if you're a developer or a business owner listening to this, what's the move? Do you start building your own skills, or do you wait for the "official" ones to come out?
The move is to look at the "friction points." If you find yourself repeatedly looking up the same local regulations, or struggling with the same Hebrew document layouts, that’s a candidate for a skill bundle. And don't reinvent the wheel. Check the Skills-IL GitHub first. If there's already a "Hebrew Date Converter" or a "Government API Wrapper," use it and contribute back to it. The more standardized these "plugs" become, the more powerful the whole ecosystem gets.
I think the takeaway for me is that "localization" is no longer a "nice to have" or a "final step" in AI development. It’s actually where the real value is created. A generic AI is a toy; an AI with an "Israeli Bureaucracy Decoder" and a "Pikud HaOref" integration is a tool for living.
It’s the difference between "artificial intelligence" and "contextual intelligence." And I think Daniel’s prompt really highlights that Israel is leading the way in showing the world how to make AI actually "live" in a specific place.
Well, I for one am looking forward to my AI agent helping me navigate my next encounter with the Israeli Tax Authority. If it can handle that, it can handle anything. Before we wrap up, we should probably give some practical pointers for people who want to explore this.
Definitely. If you’re a developer, go to GitHub and search for the Skills-IL organization. They have bundles categorized into things like finance, legal, and civic tech. If you’re more on the user side, check out agentskills-dot-co-dot-il. It’s a great directory that even includes a "Trust Score" for different skills, which is a great way to see what the community actually values and trusts.
And for the "civil defense" side, that Miklat-MCP repo is a masterclass in how to build a high-stakes, localized AI tool. It’s open source, so you can see exactly how they are pulling the data and presenting it to the agent. It’s a great blueprint for other regions in the world that might need similar systems.
It really is. And I think we'll see more of this—AI that isn't just "smart" in a general sense, but "wise" in a local sense.
"Wise in a local sense." I like that. Alright, I think we've thoroughly unpacked the Israeli AI skill ecosystem. It’s a glimpse into a future where our AI assistants aren't just global citizens, but local experts.
And that is a much more useful future.
Huge thanks to Daniel for sending this one over. It really pushed us to look at the "on the ground" reality of AI agents. If you enjoyed this dive into the weeds of localized AI, please consider leaving us a review on Apple Podcasts or Spotify—it really helps the show find new curious minds.
Thanks as always to our producer, Hilbert Flumingtop, for keeping the gears turning behind the scenes. And a big thank you to Modal for providing the GPU credits that power the generation of this show.
We'll be back next time with another weird prompt to explore. This has been My Weird Prompts.
See you then.
