The Rights Advisor AI
Last updated: November 20, 2026
What it is
The Rights Advisor is the one AI tool on this site that readers use directly. It is a chat service that answers questions about civil rights, voting rights and civic action. Claude, an AI model from Anthropic, writes the answers. Our own knowledge base supplies the facts that Claude is asked to ground them in.
It gives general information, not legal advice. It is not a lawyer, and it can be wrong.
This article describes the Rights Advisor service that runs on our backend. It is the service behind the embeddable Rights Advisor widget and our apps. The small floating chat button on the website is a separate component, built on a WordPress plugin. It also answers from our knowledge base, and it is covered in How Publishing and the Chatbot’s Index Stay in Sync and How Citizen Analyst Uses AI.
When it runs
Only when someone sends a message. There is no schedule. Each request is handled on its own, and the service does not remember earlier requests. The conversation is sent along with each new message.
What it does
- Takes the message and the mode. The service accepts four modes, described below. It rejects a mode or persona it does not recognize, an empty conversation, and a conversation whose last turn is not from the user.
- Builds the instructions for Claude. These are written in layers: the mode’s base instructions first, then any selected persona context, then retrieved knowledge base excerpts, then expert skills, then a pinned document if one was requested.
- Retrieves from our own knowledge base. The service keeps an index of the knowledge base in memory and scores documents against the question by title, summary and text overlap. The two best matches are added to Claude’s instructions as excerpts of up to about 1,200 characters each, so each one is cut short. It prefers documents from different categories. The reply lists the three best-matching documents so you can read them yourself.
- Adds expert skills. A separate step looks at the question and picks up to five of our skill files to add as method guidance. Skills are covered in Skill Router and Investigation API.
- Asks Claude and returns the answer. Answers can be streamed as they are written. If Claude declines to answer, the service reports that as an error. It does not return a blank or half-finished answer as if it were complete.
The four modes
- Rights Advisor (default). Explains constitutional rights, voting rights and state voting information. Walks through “what do I do if” situations, such as an encounter with immigration agents, being denied the right to vote, a protest stop or a workplace raid. Points to legal resources and hotlines.
- Document Generator. Drafts civic education materials for a chosen audience: a flyer, brief, talking points, letter, social post or press release. It matches the tone to the audience, uses plain language, and includes legal citations and resources where relevant.
- Conversation Simulator. A training tool. It role-plays how a type of civic actor might respond, so advocates and trainers can rehearse real conversations. After each simulated reply it adds a short debrief note explaining the communication dynamics. The available personas are a senator’s staff advisor, an organizer, a governor, a CEO, a journalist, law enforcement and a set of population personas. The persona context comes from files in our knowledge base.
- Strategy Advisor. Helps with organizing and advocacy strategy: coalition building, voter registration and turnout, rapid-response planning, media and narrative, and nonviolent civil resistance methods.
Skills and models it uses
- Models. By default, answers come from a top-tier Claude model, and a smaller, faster Claude model picks the skills. The deployed settings can differ from those defaults.
- Skills. The service reads the skill index and adds short method excerpts to the prompt. Examples of skills it may draw on include Immigration Know Your Rights Expert, Voting Rights Act Expert and Civil Resistance Theory. Which ones appear depends on the question.
- Knowledge base. Our own published articles are the source for retrieval. Which documents are left out is described under the rules below.
Rules and gates it obeys
The instructions to Claude say that it must always:
- cite specific laws, amendments and court cases where relevant;
- recommend consulting an attorney for a specific legal situation;
- emphasize nonviolent, legal approaches;
- put the user’s safety and dignity first;
- ground answers in the knowledge base context provided.
The retrieval layer has its own gates, written into the code:
- Non-public statuses are left out. Documents whose status is draft, review, research, archived, pending, unpublished, retracted or quarantined are not added to the index. A document with no status at all is indexed.
- Internal material is left out. A document marked for internal audience is not indexed. So are staging folders, raw clippings from other publishers and files at the top level of the vault.
- A short list of sensitive documents is held back unless the caller’s token has an analyst or operator role.
These are written prompts and retrieval filters. They shape and limit what Claude is given, but they do not guarantee a correct answer. A language model can still make mistakes.
For the wider set of AI guardrails, see AI and Research Runtime Guardrails. How an edit to a published article reaches the retrieval index is in How Publishing and the Chatbot’s Index Stay in Sync.
What it may never do
The instructions to Claude say it must never:
- give advice that could facilitate illegal activity;
- make assumptions about a user’s identity or citizenship status;
- store or share personal information a user gives it.
The Conversation Simulator is told never to simulate content that incites illegal activity or genuine harm. The Document Generator is told to emphasize nonviolent, legal civic action.
We also say what the service is not: it is not a lawyer, and it does not give legal advice.
Privacy and access
- Zero-PII design. The service does not ask who you are. Direct access uses a single-use invite code that is exchanged for a token valid for four hours. The token carries only an issuer, an issue time, an expiry and a role. It does not carry a name or an email address. The code is stored only as a hash and is deleted when used.
- No cookies, no stored conversations. The service does not use cookies. The code does not write your messages to a database. The request log records only the request’s method, path, result code and duration, never what you typed.
- The key stays on the server. The key that lets us use Claude never goes to your browser or a mobile app.
- Your messages do reach Anthropic. Our server sends them to Anthropic, which runs Claude, along with our instructions and excerpts. How Citizen Analyst Uses AI says what leaves our servers. Do not type your name, address, immigration status or case numbers into any chat.
- Whether a deployment requires a token is a configuration setting, and we make no claim beyond the design described here.
Where a human decides
- What it can know. Editors decide what is published and what status each document carries. Drafts are kept out of the index.
- What it is told. The base instructions and the skill index are changed only by editors and engineers through code review. Editors push a new skill index to the live service deliberately, so it can lag behind the repository.
- What you do with an answer. The advisor shows its sources so you can check them. For anything that matters, read the cited documents and primary sources, and ask a lawyer.
No person reads or approves individual chat answers as they are produced.
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.
