チームみらいの社会実験「しゃべれるマニフェスト」から得られた知見
https://gyazo.com/7770dce5b3c9d391fbca00411a22029b
nishio I wrote an article summarizing insights from the Broad Listening project run by Anno Takahiro’s political party, Team Mirai. #Pluarlity nishio 1/ New article: What we learned from Team Mirai’s “Talking Manifesto,” a large-scale Broad Listening experiment during Japan’s 2025 election. Thread nishio 2/ In May 2025, AI engineer Takahiro Anno founded Team Mirai and won a seat in the Upper House (≈2.6% vote share). That secured national party status and stable public funding for digital public goods for 6 years. nishio 3/ True to its pledge, Team Mirai formed an engineering team to build digital public goods—an innovation in how we finance “the commons,” echoing dialogues seen at Funding the Commons and ideas from Audrey Tang & Glen Weyl’s “Plurality." #Plurality #FundingTheCommons nishio 4/ Context: In the 2024 Tokyo gubernatorial race, Anno popularized Broad Listening—using AI to collect, structure, and reflect diverse public input. He placed 5th (154,638 votes; 2.3%)—remarkable for an unaffiliated newcomer. nishio 5/ 2024 introduced 3 connected projects: ① AI Town Hall (24/7 Q&A, even by phone)
② Broad Listening visualization
③ Open-source policy improvement on GitHub
—AI referenced the manifest on GitHub; insights were visualized at scale.
nishio 6/ During the 2024 campaign window (Jun 21–Jul 6): 232 problem reports, 104 change proposals, 85 adopted; the AI answered ~8,600 questions using the latest policy data. This volume would’ve been hard without these systems. nishio 7/ 2025 upgrade: “Talking Manifesto.” Voters chat with an AI interviewer that helps them articulate concerns → the AI files a GitHub pull request (PR) on their behalf. No GitHub skills required.
nishio 8/ Impact landed immediately. Within days, >10× more proposals than the governor race. Ultimately, 8.5k+ policy suggestions arrived—an unprecedented PR inflow for a political organization. nishio 9/ Challenge 1 —Throughput: I built a daily GitHub-Actions pipeline to pull PR data and auto-classify by target files for rapid triage and debate with LLMs. It hit API rate limits at 1k/hr, so we moved to diff-only, incremental updates. nishio 10/ Challenge 2 —Triage UX: Jun Ito’s auto-labeling let staff filter PRs by domain (health, education, etc.), routing to the right specialists directly in GitHub’s UI. nishio 11/ Challenge 3 —Ops at scale: With http://
Devin.ai, we demoed Slack-native commands to query all proposals and auto-generate reports in minutes—useful when the dataset ballooned.
nishio 12/ Unexpected UX pitfall: Many people chatted on the index (README) page and submitted proposals there—asking about or changing content that actually lived on subpages (e.g., “Health”). The UI nudged them to the wrong place. nishio 13/ Fix to consider next time: If a user types on the index page, search across sections before answering; if they propose edits there, redirect them to the most relevant section first. (Hard to back-fit mid-campaign, but key learning.) nishio 14/ Visualization: We used our upgraded Broad Listening tool “Kouchou-AI” (born from “Talk to the City”) to map proposals and spot patterns. Server-side analysis produced shareable URLs so policy & comms teams could act fast. #TalkToTheCity #KouchouAI nishio 15/ Re-clustering effect: Observing real voter input led us to add a new “Welfare” category and reassign many items. Broad Listening doesn’t just collect opinions—it reshapes the taxonomy of policy. nishio 16/ We then extracted problem statements from proposals and built a hierarchical map (e.g., “Inclusive education & welfare,” “Transparent, composable policymaking”). Each node links back to its source PR. nishio 17/ Biggest structural pain: PR conflicts. Multiple edits to the same lines can’t be merged simultaneously. Staff had to read conflicting edits and synthesize fresh ones—a heroic but unsustainable approach. Next time: collect problems & goals, not line-edits. 2025-08-17
締切がないからついつい先延ばしになってしまう
チームみらいの行った社会実験の成果をまとめてNoteにしたい
ここに書きかけたけど、公開前にレビューしてもらうべきだよなと思ったので後で。