Israel is heading into another election, and I have been doing the thing I suspect a lot of people here do at this point in the cycle: reading party platforms and finding almost nothing in them (I got through four before giving up). Not disagreeable things. Nothing. A page of adjectives, a photograph, a list of values that no one could object to because no one could act on them either.
The honest response available to me is a blank ballot, which I have thought about more than once. It's a gesture that says nothing here is worth my vote, and it has the specific weakness of also saying nothing about what would be.
So I tried a different gesture. Over the past few evenings I've been building a political platform for a party that does not exist (this is, I accept, a slightly unusual hobby).
It's called MK-Claude, and it's readable as a site at danielrosehill.github.io/MK-Claude. MK, as in Member of Knesset — a title Claude does not hold, is not eligible to hold, and would comfortably lose its deposit pursuing.
An AI-assisted political platform experiment for Israel — a detailed, researched manifesto for a party that doesn't exist
The actual question
The question I'm interested in isn't "can an AI write policy?" That one's boring, and the answer is a qualified yes in the same way that a very fast research assistant can write policy: it depends entirely on what you make it prove.
The question is: what would a political platform look like if somebody built one properly? Not longer. Not more radical. Just held to the standard that any think tank, any consultancy, any competent civil servant would be held to (a low bar in theory, rarely cleared in practice): every number sourced, every proposal costed, every piece of legislation cited by name and section, and every proposal naming the people it makes worse off.
That last one is the interesting constraint. Platforms don't fail because they're insufficiently ambitious. They fail because they're written as though policy has no losers, which is the tell that nobody expects to implement them. A serious document says who pays.
How it works: four stages, each with a gate
Policy moves through a pipeline, and an area can't advance until the current stage clears a defined bar.
It starts with testimony. I dictate a first-person account of something I have actually experienced (the rental market, the commute, the noise), which then gets cleaned up, dated, and committed. The raw dictations live outside the published tree on purpose, so nothing half-formed gets published by accident.
Then an evidence base: five documents per area. What the problem is once you quantify it, what Israeli law already says and why it isn't working, how three or more other countries handle the same thing (including at least one instructive failure, on the grounds that countries which got it wrong are more useful than the ones that got it right), what the statistics show, and what Israel has already tried and why that attempt died.
Then a policy paper: a proper think-tank product, typeset and downloadable (Typst, if you're curious), with costed proposals and a trade-offs section. Then, eventually, those papers roll up into a program for government.
The gates are the point. An area sitting at stage two cannot have a policy paper written about it, no matter how strongly I feel about it (and on some of these I feel quite strongly). Feeling is stage one. It does not get promoted by enthusiasm.
The rules that make it not-just-vibes
The failure mode of AI-assisted research is not laziness. It's a confident number with nothing behind it (a figure that reads exactly like a real statistic and simply is not one). Every rule in this project exists to catch that.
No invented numbers. A statistic without a source is tagged UNVERIFIED, and that tag blocks the area from advancing. It's not a note to self; it's a gate.
A fixed source hierarchy: the Central Bureau of Statistics first, then the Bank of Israel, the Knesset Research and Information Center, the State Comptroller, the OECD, peer-reviewed work, and journalism last and flagged as such.
Legislation is cited by official name, year, amendment number and the specific sections relied on. Not "the housing law".
Testimony is evidence of experience, never of prevalence. My bad experience of renting is not data about renting.
Every proposal names its losers and answers the strongest objection to itself, rather than the most convenient one.
There's also a context layer, which turned out to be the part I'd have skipped if I were being less careful. It records the political ground truth the platform is written against: the election, the outgoing Knesset, a profile per party, and a comparison matrix per policy area. The matrix answers the question a policy paper structurally cannot ask about itself — is this proposal actually needed, actually distinctive, and actually passable? Several ideas I liked turned out to be already law, or already promised by four other parties, or dead on arrival with a blocking bloc.
Our own column in that matrix gets rated on the same scale as everybody else's. This is harder than it sounds and I recommend it to anyone convinced their ideas are better than the available options.
One finding, and the caveats that come with it
The area furthest along is the rental market, and it produced a number I keep coming back to.
The CBS sub-index covering the costs of entering a tenancy (brokerage, contract, insurance, the overheads of moving rather than of living somewhere) rose 73.4% over the decade to November 2025. General inflation over the same period was 17.9%. Four times faster.
Which matches the testimony almost too neatly, so it's worth stating what the figure does not say. The sub-index bundles brokerage with contract and insurance costs, and CBS doesn't publicly decompose it, so I can't attribute the rise to broker fees alone. Its weight in the headline CPI is small, so this is a claim about the cost of moving and not about inflation generally. And it prices those costs; it says nothing about how often households actually pay them, which is a separate and currently unmeasured thing.
The same research file records that around 80% of Israeli renters reported satisfaction with their dwelling in 2022, with no notable gap between younger and older renters. That is inconvenient for the register of my own testimony, and it is sitting there in the evidence base rather than in a drawer. The reconciliation I find honest is that 20% of roughly 815,000 households is still about 160,000 households in difficulty. That is a serious problem at a large scale, and it is not the universal crisis that annoyed people (me, for instance) tend to describe.
The plugin, because of course there's a plugin
The pipeline is automated by a Claude Code plugin that ships inside the same repository: a command per stage, three skills encoding the evidence standards, and a bundled MCP server giving Claude direct tool access to CBS index data, with no API key needed (a lower barrier to entry than most statistical agencies manage).
/plugin marketplace add danielrosehill/MK-Claude
/plugin install policy-development-assistant@mk-claudeI'd suggest the plugin is the more reusable half of this. The politics are mine and you're welcome to disagree with all of them, but a pipeline that forces sourcing, comparative research, and a trade-offs section before it will let you write a conclusion is not specific to Israel, or to policy, or to me.
Where it actually stands, which is early
This is an experiment and a work in progress, and I want to be precise about how much of a work in progress. Six policy areas: the rental market, public transport, political accountability, civil preparedness and the diaspora-facing state, open government data, and environmental quality. All six have testimony. One is at two research documents out of five, another at three. Nothing has reached the policy-paper stage. There is no program for government yet; the file exists, and it is a structure rather than a manifesto.
The task list is published, which was a deliberate choice and is occasionally uncomfortable. It tracks verification debt in public, including a legend entry for claims the project is not currently entitled to make. Right now that includes a central claim of the rental paper, which rests on an unverified negative I haven't nailed down yet. It's marked as such, in public, rather than smoothed over.
A platform that claims to be evidence-based is only worth its weakest claim. Publishing while unfinished, with the unchecked numbers visible as unchecked, seemed like the only version of this that isn't quietly doing the thing I'm complaining about.
What it isn't
It isn't a party, and I'm not running for anything. It isn't an endorsement of any existing party or bloc. And it isn't an argument that AI should make policy — the human sets the values and the priorities and owns every error; the model does research, drafting, and the useful work of arguing back when a proposal is weak.
Think of it as civic engagement by other means. Instead of casting a blank ballot, writing down (carefully, with sources, and with the trade-offs admitted) what a platform worth voting for would actually contain.
The repository is public and takes issues and pull requests. Disagreement is the most useful thing anyone could send me, particularly if it arrives with a citation attached (none so far, which I am choosing not to interpret).
