Updated 27 October 2026: rewritten to be shorter and easier to read on a phone, with diagrams and a step-by-step guide to doing this yourself.
On the eve of Israel's 2026 election I recorded a 42-minute voice note while looking out of a window at the Knesset (literally: I could see the building), and asked Claude to turn it into a vote decision.
This post is how that worked, what came out of it, and how you can do the same thing with a phone and an AI chat app. The whole project is public in Israel-Election-2026-Vote-Decision on GitHub.
GitHub Repository danielrosehill/Israel-Election-2026-Vote-Decision Working out who to vote for in the 27 October 2026 Knesset election, starting from a voice note of my views TypstUpdated Oct 2026 View on GitHubThe short version
My top three: Together (רק), The Democrats (אמת) and Yisrael Beiteinu (ל).
It's stable: that order held on everyday issues alone, at a 70/30 weighting and at 50/50.
The headline finding: no list on the ballot has a real tenant-protection policy, my top issue. The best score any list got on it was 3 out of 5.
It's all public: the audio, transcripts, rubric, evidence and every score.
Why I did this
I've lived in Israel for about ten years, having moved here from Ireland. By this election I'd reached the point where even reading about the parties felt like work.
The question I actually want answered (what will you do that makes my life better?) is the one the coverage almost never touches.
So I used a workflow I already rely on for things like big purchases: talk for as long as it takes, then let AI turn the rambling into a structured research task. This time I made it public, because "how do I vote?" is about as personal a research question as there is.
How it worked
Voice note. 42 minutes, recorded on 26 October 2026. Mostly about rent, which I called "the biggest issue in my life, and it's not even close" (07:25).
Transcription. AssemblyAI. The raw transcript is kept unedited, next to two cleaned versions.
Rubric and weights. Claude Opus 5.5 designed the method on 27 October 2026: 15 criteria, each citing the timestamp where I said it. It also set the weights, including the 70/30 split, from how I talked (rent first and loudest). I didn't give it any numbers.
Two rules from the note. Lists running Netanyahu for prime minister ("the one person I'm for sure not voting for", 06:48) are ranked but flagged as excluded. Lists polling under the 3.25% threshold are named separately.
Research. Four agents gathered sourced evidence on all 37 lists (38 were filed; Blue and White withdrew on 19 October), in Hebrew and English. Evidence only, no scores.
Scoring. A separate agent scored every list on all 15 criteria, with a written reason for each score. (An agent that marks its own homework is not one I'd trust.)
Ranking and outputs. A short Python script computed the rankings and checked them against the polls. The results went into an end-to-end PDF and an interactive Hugging Face Space.
What I care about
The everyday issues ("Stream A") are each scored 0 to 5 and weighted:
| Issue | Weight |
|---|---|
| Tenant protection and rental reform | 30 |
| Housing affordability for residents | 20 |
| Cost of living and competition | 20 |
| A working state: bureaucracy, services, emergencies | 15 |
| Israel's global reputation | 10 |
| Programme quality and accountability | 5 |
The national issues ("Stream B") are nine, weighted equally, and scored by how closely a list matches my position: Haredi enlistment, the judicial overhaul, Netanyahu, the far right in government, the Palestinian question, security and alliances, religion and state, an October 7 commission, and diaspora and hasbara spending.
The combined score is 70% everyday issues and 30% national issues. Here's how that adds up:
One design choice matters a lot: a list with no stated policy on an everyday issue scores zero on it. If it isn't written down, for my purposes it doesn't exist.
The result
Together (רק): 71.5. The best everyday-issues offer: the most detailed cost-of-living plan, 100,000 long-term rental units, and the only measure aimed at lease length.
The Democrats (אמת): 63.5. Closest to me on almost every national issue, with the best-documented programme (a 16-page platform PDF). Leans on price controls rather than competition.
Yisrael Beiteinu (ל): 54.0. Concrete rental-supply and mortgage numbers, strong on competition. Nothing on tenant protection, and it leads with the Haredi draft.
Joint List (ודם): 46.8. The only rent cap of any major list (20% of household income). Far from me on trade and security.
Yashar (דרך): 46.5. Polls second overall and fits me on national issues, but offers one undetailed phrase on renters.
Lists running Netanyahu for prime minister, or pledged to recommend only him (Likud, Shas, Otzma Yehudit, RZP–Zehut, Noam, Amcha Yisrael) were excluded by my stated choice. The best match below the threshold was the Reservists and the Economic Party.
No list has a real tenant-protection policy
The best score was 3 out of 5: Together, for tax incentives on leases of five years or more. Nobody proposes:
limits on no-cause or retaliatory eviction;
enforcing the repair duties that already exist;
enforcing the ban on agent fees (already the law, and universally ignored).
It's what I suspected the night before. It was still true on election day.
Try this yourself: a guide for non-developers
You don't need code, agents or a GitHub repo. A phone and an AI chat app with web search (Claude, ChatGPT or Gemini) is enough. Allow an hour or two.
Record a voice note. Use your phone's voice recorder. Talk for 10 to 40 minutes about what you care about, what annoys you, and your red lines (anyone you'd never vote for). Don't tidy it up.
Transcribe it. Most AI chat apps accept an audio file or have voice input. Or use your phone's own transcription (many recorder apps have it). Skim the text and fix misheard names.
Turn it into criteria. Start a new chat and paste in the prompt below, followed by your transcript. Read the criteria and weights it proposes, and correct them before going on.
Research every list. Reply "confirmed" and let it search the web for every list on the ballot, citing a source for each claim. In Israel, the Central Elections Committee publishes the full list.
Score and rank. It should score each list per criterion with a one-line reason, then give you a ranked table.
Check it yourself. Open the sources behind your top few. If a weight feels wrong, change it and ask for a re-rank.
Here's the prompt. Copy it as is and fill in the two bits in square brackets:
I'm deciding how to vote in the upcoming election in [country]. Below is a transcript of a voice note where I talk about what I care about, what annoys me, and my red lines.
Step 1. Turn it into a scoring rubric: 6 to 15 criteria, each with a weight (the weights add up to 100), my position in one line, and the quote from my note it comes from. Treat anyone I said I'd never vote for as a red line, not a criterion. Show me the rubric and stop. Don't continue until I confirm or change it.
Step 2 (after I confirm). Search the web for every party or list on the ballot. Use official and party sources first, then reputable press. For each list, summarise its position on each criterion and cite a source with a link. If you can't find a position, write "no evidence" rather than guessing.
Step 3. Score each list 0 to 5 on each criterion, with a one-line reason for each score. Compute a weighted total out of 100 and show a ranked table. Flag any list that breaks one of my red lines, and any list polling below the electoral threshold.
Step 4. Tell me where the evidence was thin and which scores I should double-check myself.
Transcript: [paste your transcript here]
If the chat starts losing track, run steps 2 and 3 in batches of five or ten lists.
Use with caution
This is not advice on how to vote. It's a way of checking the parties against your own priorities. The decision is yours.
AI makes mistakes, and can be confidently wrong. Check the claims that decide your top few against their sources.
Check ballot letters at the polling station. Don't rely on a chatbot for them.
It's only as good as what you said. Leave out an issue and it won't count.
Treat it as a second opinion, not an instruction.
Mine had limits too. The 22 minor lists got only two to five searches each, so a small list with a great platform no one indexed would have been penalised. And the weights were the model's inference, not my choice.
Why I think this is empowering for voters
An election usually hands you a pile of manifestos (when they exist at all), slogans and coalition arithmetic. This flips it around: you start from your own priorities, and the parties have to answer to them.
It also shows you exactly where the parties are silent. For me that was rent.
A rant, it turns out, is a perfectly good research brief. And if any party is reading: the bar for my vote on rent is a PDF. Nobody cleared it this time.
Everything is public
The end-to-end report (PDF): the whole chain in one document, from voice note to recommendation. Start here.
The interactive version on Hugging Face: move the weights and watch the ranking change.
The project repo: audio, transcripts, rubric, evidence and scores.
The candidate-lists dataset: the Central Elections Committee lists the research started from.
Israeli-Elections-2026: my earlier project, which asked what will happen rather than who to vote for. I wrote it up in forecasting Israel's election with decision models.
