About MeetingWatch

Every meeting goes through the same fixed set of questions — the same ones, for every board and town. Without giving away the recipe, that set asks the analysis to:

• Summarize what the board discussed and decided.
• Capture what the public said during comment, and the positions they took.
• Record votes, and flag where members disagreed or a vote was split.
• Note topics that came up but weren’t on the agenda.
• Compare the discussion against the official minutes.
• Where members are known, track each one’s votes and stances — marking what was explicitly recorded versus inferred from context.

People reasonably want to know what we’re telling the AI. We won’t publish our prompts in full — the exact wording and the structure of what we ask for is the engineering behind the site. But the rules that keep a report grounded and fair aren’t a secret. Here are real, word-for-word instructions the analysis runs under:

“Be factual, concise, and neutral. Focus on decisions, votes, and action items. Do not editorialize or add opinions.”

“Base your summary ONLY on what the [record] states. Do not invent detail, timestamps, or vote tallies not present in the text. Where the [record] is silent, omit rather than guess.”

“If a vote was split but the record does not name who the dissenter(s) were, you MUST NOT guess that a member voted ‘yes.’ You do not know who dissented.”

Instructions like these are why a report says “unclear” instead of inventing a number, and why it won’t pin a vote on someone the record doesn’t name.

The short summaries we share to social media run under the same kind of rules — written to stay grounded in the record and to resist the easy gotcha:

“Write in plain language. Never sensationalize, but never sugarcoat either.”

“Be specific — name the actual issue, the date, and what was decided. Never be vague.”

“Do NOT claim information is ‘missing from the record’ … Only raise transparency concerns about information that should have been disclosed at the meeting itself.”

We also ask these summaries to notice when a decision looks driven by politics rather than evidence — but the instruction never names a party, a side, or an ideology. It asks the same question of every board, in whichever direction the politics run:

“[Flag] decisions where the board appeared to prioritize political or ideological positions over evidence, student outcomes, or fiscal responsibility.”

Because that test is symmetric — it points at ⁠any ideology overriding the facts, not one in particular — flagging it isn’t taking a side. It’s the opposite: holding every board to the same evidence-first standard, whoever is sitting on it.

The part we keep to ourselves is the rest of each prompt — how the AI is asked to organize and structure what it finds. The rules that make it fair are above; the rules that make it ⁠work are ours.

Because that question set is identical everywhere, no meeting gets a softer or a harder reading than another. A few rules keep reports grounded in the record, not opinion:

• Nothing is cherry-picked. We don’t choose which meetings to cover, or which moments to amplify, based on the topic or the people involved. Every public meeting with a recording gets a report, read start to finish.
• Recorded vs. inferred is labeled. A clearly recorded vote or direct quote is marked differently from something inferred from context — so you always know how solid a claim is.
• We don’t take sides. Reports describe what happened; they don’t endorse positions, candidates, or outcomes. When a meeting is marked spirited or heated, that reflects what was actually said — not a view on who was right.
• Our posts follow the same rule. Short social summaries may flag ⁠process problems — overdue minutes, a decision made with little public discussion — but every claim stays grounded in the record. Our accountability is about open government, not which side is right.

We’re upfront about the edges:

• It doesn’t change the words. Transcription can mishear names; we fix those against a fixed list of real local names, and nothing more. It can’t rewrite meaning or invent anything — the substance stays exactly as spoken.
• It reads the words, not the room. Working from the transcript means no vocal tone, no body language, no off-camera history. That’s deliberate — guessing at mood or motive is where bias starts — so the interpretation is left to you.
• It’s AI, and imperfect. Analysis can contain errors, and speaker labels can be wrong when people don’t identify themselves. Which is the whole reason every report links back to the source.

We want reports to be right. A correction is a factual error — a wrong vote tally, a comment attributed to the wrong person, a wrong date or name. Disagreeing with how something was emphasized or framed isn’t a correction; the same fixed process runs on every meeting, and a fair reading of the record won’t always match everyone’s preferred version of events.

Report a factual error and we’ll review it within 3 business days. When a correction changes the substance of a report, we note it on that report’s page.

This is the part that matters most. Every report links straight to its sources — the original video, agenda, and minutes. If something looks off, go watch that moment yourself: the source records are always the authority, and our report is a guide to them, not a replacement. When the two disagree, the record wins. Spotted an error? Tell us — we read every note.

MeetingWatch is built by two people who met over 30 years ago as college roommates at Tufts University. In 2004 they co-founded a software company — headquartered in Lexington, Massachusetts and still in business today — that builds enterprise content- and document-management systems and research-delivery portals for large organizations across professional services, manufacturing, publishing, and the public sector, among other industries.

MeetingWatch is a passion project, and an AI one. We’re AI-natives in our main business, and we kept coming back to a simple conviction: that AI should be used for the good of humanity. It can now gather, read, and make sense of public information at a scale — and in a way — that simply wasn’t possible before, and local government struck us as a place where that could make an enormous difference. The records have always been open; what was missing was a practical way for residents to actually follow them. This grew out of an honest curiosity — a wish for unfiltered, unbiased information about what’s happening in the rooms where local decisions get made — and a belief that the tools we’ve spent more than two decades building for large institutions could finally put that within everyone’s reach.

Running this site costs money, and we may someday support it with advertising, sponsorships, or paid features. None of that can change what a report says — because there’s no editorial function to influence. No person decides what goes into a report or which meetings get covered, so an advertiser or sponsor has no lever to pull. And the underlying records are public: you can always get them from the town directly, so our reporting stays checkable no matter how the site is funded.

All source content — meeting videos, agendas, and minutes — originates from public government meetings and records made available under state open-meetings and public-records laws. MeetingWatch provides transformative analysis of these records for civic accountability purposes, and is not affiliated with any municipality or government body. Original materials remain the property of their respective authors. If you believe any content has been published in error, contact us and we’ll address it promptly.

For the details, see our Privacy Policy and Terms of Service.