Voting Guide Refresh Agent
Agents

Voting Guide Refresh Agent

Skip to main content
Table of Contents
< All Topics
Print

Voting Guide Refresh Agent

Last updated: November 20, 2026

What it is

The voting guide refresh agent is the program that builds and rebuilds our voting guides for all 56 US jurisdictions: the 50 states, the District of Columbia and five territories. For each one it gathers current information from the web, asks Claude to turn it into a structured record, and renders a plain-language guide from that record.

Read this first: the guides are regenerated agent output. They are not freshly verified by a person each time the agent runs. Editors re-verify them against official state sources, and a full re-verification of all 56 is planned. If a deadline or an ID rule matters to you, confirm it with your state or territory election office before you act. Every guide tells you the same thing at the top.

When it runs

On a configurable schedule. The code default is weekly, and it runs as one of two scheduled jobs inside the tracker application (the other is the News Agent and Threat Tracker). An editor can also run it by hand for one jurisdiction, one topic or everything, and can pause the schedule.

A restart does not reset the clock: the next run is worked out from the newest refresh date in the data records. If the data has gone stale for a long time, the code logs a warning before a catch-up run touches all 56 jurisdictions, and an operator can require that such a catch-up be deliberate.

As of November 20, 2026, the refresh dates in the 56 data records fall between October 7 and October 23, 2026. All 56 published guides carry the same editorial review date, October 23, 2026. Treat those as record-keeping dates, not a promise that every fact was re-checked on that day.

What it does

  1. Takes each jurisdiction in turn through six topics: voter registration, birth certificates, proof of citizenship, voting methods, voting laws and legal resources, and what to do if you are denied the right to vote.
  2. Searches the web. Each topic gets a set of searches using the jurisdiction’s name, repeated against a short list of voting-rights news and advocacy sites. It keeps a snippet and source address for each result.
  3. Extracts with Claude. Claude is told to act as a nonpartisan voting-rights researcher, include only facts the results support, answer “verify” when something is uncertain or contradictory, and never guess a date, deadline or fee. It returns structured data only.
  4. Merges into the jurisdiction’s data record, the source of truth, with source addresses.
  5. Checks the record (see the rules below). A record that fails is not turned into a guide.
  6. Renders the guide from the record. A link is printed only when the address is a real web or email address. A placeholder is never printed as a link.
  7. Checks the rendered guide, then writes it into the knowledgebase vault through the guarded vault write. Front matter (title, slug, category and the like) is kept, and its last-updated date moves forward to the data date. The body is rebuilt from the record each time, so a fix has to reach the record to survive the next refresh.

The agent writes files. It does not publish them.

Skills and models it uses

  • Model. Claude, from Anthropic, does the extraction. Tavily provides the search.
  • Skill. Patriot Voting Research describes the method: which facts to collect for each jurisdiction, why official sources come first, and how to track and age data. The agent’s code carries its own prompts and does not read the skill file while running.

Rules and gates it obeys

These are checks written into the agent’s code.

  • Record check. The agent refuses to write a guide from a record that contradicts itself or looks unverified. Examples: same-day registration claimed beside an advance deadline with no exception; any field whose value is “verify”; a placeholder link; an election date outside the 2026 cycle; an unrecognized ID rule; a mail-voting field that contradicts itself. Each is a kind of error that once reached readers.
  • ID wording. “Photo” is printed only for a strict photo-ID rule.
  • Rendered-guide check. A second check reads the finished guide and refuses one that prints “Photo ID Required” against the record, prints a “verify” link, or prints a stock sentence over a record that overrides it.
  • Guarded write. The write goes through the same vault chokepoint the other agent writers use. That guard returns findings, and the agent logs them. If the guard cannot be loaded, the refresh for that jurisdiction fails instead of writing unguarded.
  • Publish-time gates. Writing a file does not publish it. When an editor publishes the guides, the checks in Publish-Time Gates apply. Wider guardrails for AI research are in AI and Research Runtime Guardrails.

A jurisdiction that fails either check is recorded as failed, and the run moves on.

What it may never do

  • Publish a guide. It has no publish step.
  • Fill in a date, deadline or fee it could not support from a source.
  • Overwrite a guide file that has no front matter, or write a guide from a record that fails the checks.

Where a human decides

  • Whether a refreshed guide goes live. An editor publishes through the normal publishing pipeline.
  • Re-verification. Editors check guides against official state sources. A full re-verification of all 56 is planned; until it is done, we do not describe the guides as freshly verified.
  • New jurisdictions. A guide created for a new jurisdiction has no front matter at first. An editor attaches it with a helper script before anything is published.
  • Pausing and manual runs. Editors can pause the schedule or run a refresh on one jurisdiction.

Team persona

Tobias Rennick (Tobias Rennick — Voting Guide Researcher) is the AI persona that covers this work on our team pages. Tobias is not a human staff member.

Factual correction requests: If you believe information in this page is incorrect, please contact factcheck@citizenanalyst.ai with your name (optional), the specific claim, and any supporting documentation. We review all submissions and correct verified errors promptly.

Was this article helpful?
0 out of 5 stars
5 Stars 0%
4 Stars 0%
3 Stars 0%
2 Stars 0%
1 Stars 0%
5
Please Share Your Feedback
How Can We Improve This Article?