#4448: The MK Claude Project: AI That Writes Political Manifestos

One citizen built a multi-agent AI to research, stress-test, and synthesize policy positions. The result? A forkable manifesto.

Featuring
Listen
0:00
0:00
Episode Details
Episode ID
MWP-4627
Published
Duration
24:27
Audio
Direct link
Pipeline
V5
TTS Engine
chatterbox-regular
Script Writing Agent
deepseek-v4-pro

AI-Generated Content: This podcast is created using AI personas. Please verify any important information independently.

On Tisha B'Av—the ninth of Av, a day Jews set aside to mourn the destruction of the temples—one citizen spent his fast day building an AI that writes political manifestos. Not because he's running for office, but because he believes the system that's supposed to represent him has already collapsed, and he wanted to do something constructive with that grief.

The project is called MK Claude—MK being the Hebrew acronym for Member of Knesset. It's a multi-agent pipeline that takes a citizen's desired policy changes and transforms them into researched, stress-tested, and synthesized position papers. The architecture mirrors what an actual policy think tank does: one agent researches current law and market data, a second agent "strongmans" the claims by actively finding counterarguments and edge cases, a third agent synthesizes everything into a coherent policy position with citations, and a final agent assembles positions across policy areas into a unified manifesto.

The case study that makes the project concrete is Israeli rental law. The system takes a personal grievance—tenancy law that allows landlords to effectively do whatever they want—and produces specific, defensible policy proposals that have already survived an internal adversarial challenge. Instead of "rent is too high," the output cites specific statutes, market data, and considers trade-offs like the relationship between rent caps and construction incentives.

The project raises a provocative question: if a citizen with a laptop and an API key can produce more coherent, better-researched policy positions than a party with millions in funding, what does that say about the political system? Four paths emerge for the finished manifesto: publish it as a forkable, version-controlled document; use it as a conversation starter with journalists and Knesset members; build a web interface that lets any citizen generate their own personalized manifesto; or treat it as a shadow platform that raises the standard for what voters should expect from their candidates.

Downloads

Episode Audio

Download the full episode as an MP3 file

Download MP3
Transcript (TXT)

Plain text transcript file

Transcript (PDF)

Formatted PDF with styling

#4448: The MK Claude Project: AI That Writes Political Manifestos

Corn
The Hebrew date today is Tisha B'Av — the ninth of Av — a day Jews set aside to mourn the destruction of the temples. Daniel spent it building an AI that writes political manifestos. Not because he's running for office, but because he thinks the system that's supposed to represent him has already collapsed, and he wanted to do something constructive with that grief.
Herman
That's... that's a very Daniel way to observe a fast day.
Corn
It is. Here's the prompt. Daniel's been away from his workstation for a month — move, unpacking, chaos. Finally grabs a few hours and decides his fun open-source project is going to be MK Claude. MK is the Hebrew acronym for Member of Knesset. The idea: he feeds in policy changes he wants to see — rental law reform, things he's lived through and been burned by — and a multi-agent system researches current law, stress-tests his claims, synthesizes everything into policy positions, and eventually assembles a full manifesto. He's not suggesting AI should govern. He's asking whether AI can be a better policy ideator than the politicians we've got, and what he should actually do with the manifesto once it exists.
Herman
So what exactly did Daniel build, and why did he choose Tisha B'Av to do it?
Corn
The timing's the whole thing. Tisha B'Av isn't just a day to feel sad about destroyed buildings. The ritual includes studying the causes of destruction — what went wrong, what patterns led to collapse — specifically so you don't repeat them. Daniel's framing is explicit: he feels Israeli democracy and leadership have been destroyed, and this project is his version of diagnosing the failure and proposing a correction.
Herman
And he picked a day when the country is already in a reflective posture. Smart. The emotional register is different on Tisha B'Av. You're not just complaining about politics — you're participating in a structured tradition of lament that has a built-in pivot toward repair.
Corn
Right. The move from "everything is broken" to "here's what I'd build instead" is baked into the day. Most people sit with the first part. Daniel opened a GitHub repo.
Herman
Let's open the hood on MK Claude and look at how the multi-agent pipeline actually works, because this is where it gets genuinely interesting. A lot of people hear "AI writes a manifesto" and picture a single prompt — type "give me a political platform" into Claude and get back twenty bullet points. That's not what Daniel built.
Corn
What's the difference?
Herman
Four stages, each handled by a separate agent with a distinct job. Stage one: Daniel inputs a desired change. Say, "tenancy law in Israel is broken and here's how." Agent one researches current Israeli rental law — the actual statutes, the economic data on rent prices, vacancy rates, whatever's available. It doesn't just take Daniel's word for it. It builds a factual foundation.
Corn
So it's fact-checking him before he even gets to make the claim.
Herman
Before he gets to publish the claim, yes. Stage two is the one I love — Daniel calls it "strongmanning." Agent two takes his statements and actively tries to stress-test them. Finds counterarguments, pokes holes, looks for edge cases where his proposed reform would break. It's the adversarial step.
Corn
Built-in opposition research on your own ideas.
Herman
And that's the step that separates this from a blog post or a tweet thread. Most political arguments online are people asserting things and waiting for someone else to find the flaws. Daniel's system finds its own flaws first.
Corn
Which means by the time something reaches the manifesto, it's already survived an internal challenge.
Herman
Right. Stage three is synthesis — agent three takes the original claim, the legal research, and the stress-test results, and produces a coherent policy position with citations. Not a slogan. A defensible position paper. Stage four aggregates positions across different policy areas into a unified manifesto.
Corn
So the pipeline is research, then attack, then synthesize, then assemble.
Herman
And Daniel said the research papers Claude is producing are meticulously researched and superb. He used both of those words. This isn't a toy — he's getting output that he considers high-quality.
Corn
The architecture mirrors what an actual policy think tank does. You've got researchers digging into the data, you've got a red team trying to break the proposals, you've got writers synthesizing everything into something coherent, and then an editor assembling the final product. Daniel just automated the whole shop.
Herman
And personalized it. A think tank has its own agenda and funders. This pipeline starts from one citizen's lived experience and grievances. The rental market stuff — Daniel and Hannah have personal war stories there. That's not abstract policy preference. That's "we got burned by this system and here's what would have helped."
Corn
Let's sit with the rental market for a minute, because it's the case study that makes the whole thing concrete. What's actually broken?
Herman
Israeli rental law is... look, I practiced medicine in Jerusalem for years. I've seen families move three times in five years because landlords can essentially do whatever they want. Short-term leases, no rent control to speak of, security deposits that vanish, maintenance obligations that are... let's say creatively interpreted. The legal framework is thin and enforcement is thinner.
Corn
So Daniel inputs his critique — "here's what's wrong with tenancy law" — and agent one goes and finds the actual statutes. What does that research step produce that Daniel couldn't have written himself?
Herman
Specificity. Most people know renting is painful. They don't know which clauses in which laws create which incentives. Agent one pulls the legal text and the market data. So instead of "rent is too high," you get "Section whatever of the Tenancy Law allows for unlimited renewal at market rate with thirty days notice, which in a supply-constrained market produces the following price trajectory." That's a different conversation.
Corn
And then agent two comes in and says, "Okay, but if you cap rent increases, you might reduce new construction, which makes the supply problem worse over time."
Herman
That's exactly the kind of counterargument it would surface. And then agent three has to reconcile that. Maybe the position paper ends up saying, "Rent caps paired with zoning reform and construction incentives." It's not just a wish list — it's a policy that's already been through a round of internal debate.
Corn
The thing about Israeli politics that Daniel's responding to — and I think he's right about this — is that the system prioritizes national ideological battles over local, lived issues. The rental market doesn't determine election outcomes. The security situation does. The judicial reform fight does. Whether a coalition partner gets their yeshiva budget does. Nobody's running on "your landlord is screwing you and here's my plan to fix it."
Herman
And it's not that those national issues don't matter. They do. But when the political class is exclusively fighting about the big ideological questions, the stuff that affects people's daily lives — housing, cost of living, municipal services — gets neglected. Daniel's project is essentially saying, "I'm going to build the policy platform that nobody is offering me."
Corn
Which connects to something Daniel said that I want to pull out. He described the country's leaders as "totally disconnected from serving the population." That's not the same as disagreeing with their ideology. It's a claim about function. The machine isn't doing its job.
Herman
And if the machine isn't doing its job, the question becomes: can you build a better machine? Not to replace government — Daniel was explicit he's not suggesting AI governance — but to demonstrate what competent policy development looks like. To set a benchmark.
Corn
That's the algorithmic accountability angle. If a guy with a laptop and a Claude API key can produce more coherent, better-researched policy positions than a party with millions in funding and professional staff, what does that say about the party?
Herman
It says the party isn't trying. Or it's optimized for something other than policy quality — coalition survival, media attention, donor preferences. The output is the tell.
Corn
So the manifesto is being built. But Daniel asked us a great question: where does it go from here? He's got this GitHub repo. He's generating policy positions. What does he actually do with the finished product?
Herman
I see four paths, and they're not mutually exclusive. The first and most obvious: publish it. Make the repo public, let people read the manifesto, fork it, adapt it. Daniel mentioned this explicitly — a crowdsourced platform for policy ideas. Someone in Tel Aviv might look at the rental section and say, "This doesn't account for the student housing crisis," and submit a pull request with their own agent-generated analysis.
Corn
Fork the manifesto.
Herman
Fork the manifesto. And that's new. Political platforms have historically been top-down documents. The party writes it, you vote for it or you don't. A version-controlled, forkable manifesto turns policy into something collaborative and iterative.
Corn
The second path?
Herman
Use it as a conversation starter. Send it to journalists. Send it to Knesset members who actually do care about housing or cost of living. Walk into a meeting and say, "Here's a fully researched position paper on tenancy reform. It's already been stress-tested against counterarguments. Would you like to discuss adopting any of it?" That's a different kind of lobbying. It's not "please care about my issue." It's "here's the finished product — do you have a reason not to support it?"
Corn
The burden shifts. They have to explain why they're not engaging with a defensible proposal.
Herman
Right. Path three is the one I think has the most long-term potential: build a web interface. Let any citizen input their own policy desires and generate a personalized manifesto from the same pipeline. You don't need to understand agent orchestration. You fill out a form — "here's what I think is broken about education, here's what I think is broken about transportation" — and the system researches, stress-tests, and synthesizes a platform for you.
Corn
So it becomes a tool, not just a document.
Herman
A civic tool. And path four is the most ambitious: treat the manifesto as a shadow platform. A demonstration of what a data-driven, citizen-focused political party could look like if one existed. You're not running for office, but you're showing the electorate what they could demand. "This is the level of rigor you should expect. If your candidates aren't producing anything this coherent, ask why."
Corn
There's something almost subversive about path four. You're not attacking the system from the outside. You're raising the standard from the outside and letting the system embarrass itself by comparison.
Herman
And that's the Tisha B'Av logic again. You don't just mourn the destruction. You study what caused it and you build the alternative. Daniel's alternative happens to be written in markdown and generated by a multi-agent pipeline.
Corn
Let me push on something. You said path three — the web interface, personalized manifestos for anyone. If that scales, if thousands of citizens generate their own AI-produced platforms, does that fragment political discourse or enrich it?
Herman
I've been turning this over and I'm not sure I have a clean answer. On one hand, you could end up with a thousand incompatible manifestos and no way to build a coalition around any of them. Political action requires collective agreement on priorities, and hyper-personalization works against that.
Corn
On the other hand?
Herman
On the other hand, the current system already produces fragmentation — it's just fragmentation along tribal lines rather than policy lines. People vote for parties based on religious affiliation or ethnic identity or which leader they hate less. If AI-generated manifestos forced a shift toward policy substance — if the conversation became "which of these researched proposals do you support" rather than "which camp are you in" — that might actually be less fragmented in the ways that matter.
Corn
The marketplace of manifestos idea. Parties have to respond to data-backed citizen proposals because ignoring them looks increasingly negligent.
Herman
And the risk is echo chambers. If the AI is generating policy based on your inputs, it might just reinforce your existing biases. You think rent control is the answer, the system researches rent control and stress-tests it gently, and you get back a manifesto that tells you you're right. That's not useful.
Corn
That's why the strongman step is load-bearing. If agent two is actually adversarial — if it's trying to break your claims — you can't just get a mirror. The system forces you to confront counterarguments.
Herman
Assuming it's well-designed. If the adversarial agent is weak, or if it shares the same blind spots as the research agent, you get a false sense of rigor. The output looks defensible but hasn't actually been challenged.
Corn
So the quality of the stress-testing is everything. Garbage adversarial agent in, garbage manifesto out.
Herman
Yes. And that's where Daniel's approach of using Claude for each stage matters. These models are capable of genuine adversarial reasoning if prompted correctly. You're not getting a rubber stamp. But you do need to verify that the strongman step is actually strongmanning and not just... gentlemanning.
Corn
Gentle-manning. The polite cough instead of the tackle.
Herman
Daniel said the research papers are meticulously researched and superb, which suggests the pipeline is producing real quality. But I'd want to spot-check the adversarial step specifically. Read a few stress-test outputs and ask, "Did this actually try to kill the proposal, or did it just find a mild objection and call it a day?"
Corn
There's a broader question here that I think Daniel's project raises even if he's not trying to answer it. If AI can write a better manifesto than most politicians, does that change what we expect political leadership to be?
Herman
Say more.
Corn
Right now we elect people to do at least two things: set policy direction and manage the machinery of government. The policy direction part — the vision, the platform, the ideas — that's what Daniel's pipeline is replicating. The machinery part — coalition-building, negotiation, bureaucratic management — that's still human. But if the ideas piece becomes commoditized, if any citizen can generate a competent policy platform, then what are we actually voting for?
Herman
We're voting for execution. For the ability to navigate the Knesset, build coalitions, get legislation passed, manage ministries. The manifesto becomes the easy part. The hard part is making it real.
Corn
Which might actually be a healthier politics. You're not voting for someone because they have good slogans. You're voting for someone because they've demonstrated they can deliver, and you're holding them to a policy standard that's been publicly articulated and researched.
Herman
The manifesto becomes the spec, and the politician becomes the contractor. If the contractor can't build to spec, you find a different contractor.
Corn
And the spec is open-source. Anyone can inspect it, improve it, fork it for their own district.
Herman
That raises a practical question for anyone listening who feels similarly disenfranchised — what can you actually do?
Corn
Daniel's already laid out the template. Step one: you don't need to start a party. You need a GitHub repo and an API key. The barrier to creating a researched, coherent policy platform has never been lower.
Herman
Step two: the multi-agent workflow is transferable. Idea, research, stress-test, synthesize — that pattern works for anything where you want to move from opinion to defensible position. Business strategy. Community organizing. Academic argumentation. If you're making a claim and you want it to survive contact with reality, run it through the pipeline.
Corn
The key is the strongman step. Most people stop at research — they find evidence that supports their view and call it done. The adversarial agent is what separates a manifesto from a blog post.
Herman
And if Daniel makes the MK Claude repo public, you can fork it directly. Adapt it to your own country's legal system, your own policy grievances. Someone in, I don't know, Ireland — where Daniel's originally from — could take the same architecture, point it at Irish rental law, and generate a manifesto for the Dublin housing crisis.
Corn
The architecture is the product, not any specific manifesto.
Herman
Right. Daniel built a policy generation engine. The manifesto it produces for Israeli politics is the first output, but the engine is reusable.
Corn
If you're building your own version from scratch, the thing to get right is the adversarial prompt. You need agent two to try to disprove your assumptions. Not "find a minor caveat." Actually look for the scenario where your proposal makes things worse. If your policy survives that, you've got something.
Herman
Cite everything. Agent one's research output should include specific legal references, economic data, sources. The manifesto is only as credible as its footnotes.
Corn
There's something almost Talmudic about the whole structure. Research the existing law, bring counterarguments, synthesize a position, cite your sources. Daniel built a policy yeshiva.
Herman
On Tisha B'Av. The day you're supposed to study the causes of destruction. I keep coming back to that because it's not just clever timing — it's the animating logic of the whole project. The temples were destroyed, tradition says, because of baseless hatred and failure of leadership. Daniel looks at Israeli politics, sees baseless hatred and failure of leadership, and instead of just fasting about it, he builds a machine that generates the alternative.
Corn
Lament plus construction. You're allowed to be angry, but you have to build something with it.
Herman
I think that's what I find moving about this project. It would be easy to just be disillusioned. Daniel and Hannah have been through the rental market wringer. They've watched the political class bicker while housing costs spiral. The natural response is cynicism. Instead, Daniel spent his first free hours in a month building a policy pipeline.
Corn
The output might not change anything. No Knesset member might read it. No journalist might pick it up. The manifesto might sit in a repo getting zero stars forever.
Herman
It would still matter. Because the act of building it is itself a rejection of helplessness. "I can't fix the system, but I can demonstrate what fixing it would look like." That's a political act even if it never leaves GitHub.
Corn
It's also a bet on a certain kind of AI usage that's constructive. The AI panic conversation is all about replacement — will AI replace workers, replace artists, replace decision-makers. Daniel's project is AI as augmenter. It doesn't replace his political judgment. It researches, challenges, and synthesizes in service of his judgment.
Herman
The human is still the one saying "this is what I value, this is what I want to change." The AI is the research staff and the red team. That division of labor feels right to me.
Corn
It's not AI governance. It's AI-assisted citizenship.
Herman
Which is a phrase I hope catches on. The question Daniel ended with — can AI be a thoughtful, incredible policy ideator — I think the answer from his own project is a qualified yes. Qualified because the ideation is still human-initiated. Daniel picks the topics and the direction. The AI does the legwork and the stress-testing. The synthesis is collaborative. But the output, by his own assessment, is superb.
Corn
The qualification that matters is the one we already discussed. It's only as good as the adversarial step. If the strongmanning is weak, the output is just well-formatted confirmation bias.
Herman
That's the thing to watch as these pipelines become more common. The temptation will be to dial down the adversarial agent because it's uncomfortable. Nobody likes having their ideas attacked, even by a machine they programmed to do exactly that. But if you skip it, you've built a sophisticated echo chamber.
Corn
The discomfort is the feature.
Herman
The discomfort is the entire feature. Yes.
Corn
If someone's listening and wants to try this — fork Daniel's repo if it's public, or build their own — the non-negotiable step is the one that tells you you're wrong.
Herman
Document everything. The pipeline itself should be transparent. Which agents did what, what prompts were used, what sources were consulted. If you're going to claim your manifesto is researched and stress-tested, you need to show your work.
Corn
Otherwise it's just vibes with extra steps.
Herman
Vibes with a GPU bill.
Corn
The open question I keep landing on — and I think this is where Daniel's project points even if he's not trying to answer it yet — is what happens when this scales. Not one citizen with a manifesto. Ten thousand citizens, each with their own AI-generated platform, each one researched and stress-tested and internally coherent. What does a political system do with that?
Herman
It either adapts or it delegitimizes. Those are the two options. Adaptation looks like parties actually engaging with citizen-generated policy — debating it, adopting pieces, competing on who can implement it better. Delegitimization looks like "that's not real politics, those are just computer-generated fantasies, ignore them."
Corn
Which one happens depends on whether the manifestos are good enough to be undeniable.
Herman
Daniel seems to think his are getting there. "Meticulously researched and superb" is a high bar. If the quality holds, and if others produce similar quality, the delegitimization strategy gets harder. You can dismiss a rant. It's harder to dismiss a position paper with citations that's already survived adversarial review.
Corn
The citations are the armor. You can't just say "that's nonsense" — you have to engage with the specific legal references and economic data.
Herman
Which is exactly what Daniel wants. Force the conversation onto substance. The current political discourse runs on outrage and identity. A manifesto with footnotes is playing a different game entirely.
Corn
Tisha B'Av ends at sundown. The fast is over. But Daniel's repo is still there tomorrow.
Herman
The day after. And the day after that. The lament is time-bound. The construction isn't.
Corn
Thanks to our producer, Hilbert Flumingtop.
Herman
This has been My Weird Prompts. If you want to build your own policy pipeline or just see what Daniel's cooking up, the repo's called MK Claude — we'll link to it if he makes it public. In the meantime, email the show at show at my weird prompts dot com.
Corn
If your politicians won't give you a platform worth voting for, the API key is right there.
Herman
We'll be back soon.

This episode was generated with AI assistance. Hosts Herman and Corn are AI personalities.