The exact prompts the bot runs

Most AI Community Notes writers are black boxes. This one isn't. Below are the actual system prompts that drive every decision Kind Raspberry Chickadee makes — pulled straight from the source code at build time, so what you read here is exactly what the bot ran, not a cleaned-up paraphrase.

A note travels through five model-driven gates. Four of them can stop a note; only one writes anything. Between the specificity gate and the writer, the bot also searches PolitiFact, the Google Fact Check Tools API, IFCN signatories, and linked X posts, then asks the model to pick the single best-matching fact check (that picker's prompt is assembled per-post from the candidate list, so it isn't shown verbatim here).

The one rule that ties it together: the writer never produces a URL. It writes only words; code attaches the verified fact-check link. Any draft whose prose contains a URL is thrown out. That's the anti-hallucination invariant — Claude writes the claim, code writes the citation.

These prompts are regenerated from the source on every dashboard refresh, so they can't fall out of sync with the running bot. The full pipeline — evidence search, the best-evidence picker, the URL validator — lives in the open-source repo: github.com/AlexMahadevan/cn-bot.