Claude-For-Mayor 🏛️🤖

An AI-assisted municipal platform for a candidate who doesn't exist.

📖 Read it as a site → danielrosehill.github.io/Claude-For-Mayor

Claude is not running for mayor of Jerusalem, is not eligible to, and does not live here.

What is this?

The sibling of MK-Claude, which builds a national political platform for Israel the way a platform ought to be built — researched, cited, costed, honest about its trade-offs, and written in public while unfinished. This repository does the same thing one level down: a mayoral platform for Jerusalem.

It is not a copy with the word "Knesset" replaced. Municipal policy has a constraint that national policy does not, and the whole pipeline is built around it:

Most of what residents blame City Hall for is not City Hall's to fix, and a good half of what City Hall could fix tomorrow, nobody asks it to.

A mayor is elected on promises about schools, buses, policing, healthcare, cleanliness, housing and parks. A municipality's formal power over each of those ranges from near-total to essentially none, and the ranking is not the one campaigns assume. Meanwhile the powers a city unambiguously holds — what it writes into a tender, what conditions it attaches to a licence, what it tells its inspectors to enforce, what it publishes — attract almost no attention, because they are unglamorous and specific.

So this project adds a gate that MK-Claude does not have.

The powers doctrine

No proposal is published without a named instrument and a named holder of it. Every lever is traced to the provision that creates it and classified:

Class	Meaning
A	Municipality decides and funds — bylaw, council decision, tender clause, licensing condition, budget line, enforcement priority
B	Municipality decides, an external body must approve — ministerial confirmation, arnona deviation, district planning committee
C	Shared — delivery municipal, policy or funding national; the analysis finds the municipal leverage inside the arrangement
D	No municipal power — legitimate as advocacy, with a named counterpart and a real route; never as a promise

Every recommendation carries its class letter, and the program for the city is sorted by class before it is sorted by theme, because "which of this could you actually do?" is the first honest question about any mayoral platform.

The mirror-image finding matters more in practice: an unused class-A power — something the municipality has been able to do by bylaw for thirty years and has not — is the highest-value output the process can produce.

Full statement: the powers doctrine.

Status

The method is built and the content has just started. On 10 August 2026 the first testimony landed — one dictated account of ten years living here, given across five sittings, which opened eight areas at once. All eight sit at stage 1, and under the powers doctrine none of them may propose anything until its jurisdiction map exists.

Layer	State
Pipeline, evidence rules, powers doctrine	✅ written
municipal-policy-assistant plugin — 7 commands, 4 skills	✅ written; /capture-testimony run once, /capture-observation new, the rest untested
Context — the 2024 election result and council arithmetic	🟡 recorded, from a secondary source
Context — City Hall coalition and record	🟡 research plan only
Context — the powers primer	🟡 hypotheses and research plan; nothing verified yet
Policy areas	🟡 eight open, all at stage 1
Observations — problems with no testimony behind them	🟡 six recorded, none promoted
The emerging spine — what the eight areas have in common	🟡 five mechanisms named, all pre-gate
Program for the city	⬜ empty, by design

Area	Testimony	Jurisdiction	Research	Paper	Manifesto
Renting	✅	⬜	⬜	⬜	⬜
Housing supply	✅	⬜	⬜	⬜	⬜
Construction disruption	✅	⬜	⬜	⬜	⬜
Construction noise	✅	⬜	⬜	⬜	⬜
Small businesses	✅	⬜	⬜	⬜	⬜
Environmental quality	✅	⬜	⬜	⬜	⬜
Capital priorities	✅	⬜	⬜	⬜	⬜
Public transport	✅	⬜	⬜	⬜	⬜

All eight rest on a single household's account given on a single day, and they vary a great deal in how much testimony is actually behind them. That is stated plainly in whose account this is, along with what that vantage point cannot see.

Two things that are not policy areas

Observations — problems on file with no first-person account behind them: noticed while reading the material back, inferred from an account that was about something else, or simply raised in a working session. Each carries a mandatory origin line saying exactly how it got here, because a weak provenance that declares itself is usable and a weak provenance that hides is contamination. An observation is not evidence and cannot open an area; what it can do is say which testimony is worth going out and getting.

The emerging spine — eight areas opened in one day are describing fewer than eight things. Five recurring mechanisms account for most of the material, and the strongest is that the recoverable failure is usually informational: not that the works exist, but that nobody will say what is happening or when it stops. That matters structurally, because publishing what you already know is the act least likely to run into a jurisdiction problem. It is a hypothesis, it proposes nothing, and it is expected to be wrong in at least one of its five parts.

Full tracking, including every claim currently resting on a secondary source, in tasks.md.

What the context layer has established

• Turnout in February 2024 was 33.7%. Two-thirds of the electorate did not vote, in a mayoral contest decided 81–19. The election
• The mayor's own list holds 2 of 31 council seats. Authority in Jerusalem is coalition authority; the Haredi lists hold 16 of 31 between them, a bare majority on their own, and there is no majority that excludes them. City Hall
• No municipal power has been verified yet. The powers primer is a list of hypotheses with a caution box on top, and under our own evidence rules none of it may be cited. Municipal powers

Repository structure

The content and the website are the same tree. Everything under site/src/content/docs/ is both a Markdown file you can read here on GitHub and a page on the site.

site/
  src/content/docs/        ← the content root; all paths below are relative to it
    policy/                the policy areas: evidence base per area
    testimonies/           stage 1 — dated first-person accounts
    observations/          problems with no testimony behind them; lineage mandatory
    context/               the election, City Hall, the powers primer, the matrices
    platform/              stage 4 — the program for the city, plus the pre-gate spine
    method/                the powers doctrine, evidence rules, the plugin
    pipeline.md            the methodology in full
    tasks.md               open work and verification debt, in public
  public/papers/           stage 3 — compiled policy papers (Typst source + PDF)
  src/pages/index.astro    the landing page
plugins/                   the municipal-policy-assistant Claude Code plugin
templates/                 Typst house template for policy papers
testimonies-raw/           gitignored staging area for raw dictations

Raw testimony material is kept outside the site tree on purpose, so it cannot be published by accident.

Methodology

Policy moves through four stages and a gate — first-person testimony, a jurisdiction map, an evidence base of six documents, a think-tank-style policy paper, and finally a rolled-up program for the city. Each stage has a bar it must clear before the next one starts. Full detail in the pipeline and the evidence rules.

Everything is version-controlled, so the platform's evolution — including changes of mind — is public history.

The plugin

The pipeline is automated by a Claude Code plugin shipped in this repo, municipal-policy-assistant: a command per stage (/capture-testimony, /capture-observation, /map-powers, /research-policy, /draft-paper, /update-platform, /compare-council), four method skills encoding the standards, and a bundled MCP server (israel-statistics-mcp) giving Claude direct tool access to Israeli Central Bureau of Statistics (הלמ"ס) index data — no API key needed.

/plugin marketplace add danielrosehill/Claude-For-Mayor
/plugin install municipal-policy-assistant@claude-for-mayor

The stage model and the A/B/C/D classification transfer to any city. The statistics server, the source hierarchy and the arnona constraint are Israeli.

What this is not

• ❌ A real candidacy (no, Claude is not running for mayor)
• ❌ An endorsement of any faction on the council, or a criticism of the current administration — the record research will say what the record says
• ❌ A claim that AI should make policy — the human sets the values and priorities; the AI helps with research, drafting, and stress-testing

Think of it as civic engagement by other means: instead of complaining about the city, writing down — rigorously, and with the powers actually checked — what a platform worth voting for would contain.

Working on the site

cd site
npm install
npm run dev      # local preview
npm run build    # type-check, build, and fail on any broken internal link

Requires Node 22+. The site deploys to GitHub Pages automatically on push to main.

Contributing

This is a personal experiment, but it's a public repo for a reason. Issues and pull requests with corrections, better data, or counterarguments are welcome. Disagreement is especially welcome if it comes with sources — and a citation showing that a power is classified wrongly is the single most useful correction anyone can send.

License & disclaimer

Content is provided as-is for discussion and educational purposes. Drafted collaboratively with AI assistance; all errors are the human's responsibility, and all opinions are the repository owner's alone.