Detailed Analysis
A self-described "somewhat coder" has used Claude to build and publicly launch The Influence Registry, a campaign finance transparency tool that scores all 537 sitting members of Congress on the degree to which their funding comes from special-interest money versus individual donors. The project pulls directly from FEC filings and applies a single, uniform scoring formula across every member, explicitly avoiding manual adjustments that could introduce bias or undermine credibility. Each congressional profile includes a donation breakdown, top financial supporters and opponents from outside spending groups, small-dollar donation share, net worth, and tenure — all traceable back to the original public filings. The creator has also open-sourced the project on GitHub, inviting scrutiny of both the underlying data and the methodology.
The technical details offered in the piece are notable for what they reveal about how a non-professional developer can use an AI coding assistant to manage a nontrivial engineering problem. The build consists of a Python pipeline that processes FEC data into scores, paired with a single large HTML file — reportedly around 5,000 lines — that powers the front-end application. Editing a monolithic file of that size from a phone introduces real risk, and the creator describes an incident where a greedy find-and-replace operation corrupted the markup. In response, the workflow evolved to use exact-string replacement plus a set of post-edit integrity checks (element counts, div balance, JSON validation) to catch corruption before it ships. This kind of iterative, guardrail-building process — arrived at through trial and error alongside Claude — reflects a broader pattern among non-engineers using AI assistants: the human supplies domain judgment and quality thresholds, while the AI handles syntax, boilerplate, and repetitive execution, with the pair jointly discovering the failure modes that necessitate safety checks.
This matters beyond the specifics of one campaign finance tracker because it's a concrete illustration of how generative AI is lowering the barrier to building functional, data-driven civic tools. Government transparency data — FEC filings, lobbying disclosures, voting records — has long existed in public form but has remained practically inaccessible to ordinary citizens because parsing and structuring it required programming skills most people don't have. Tools like this promise to convert "the data exists somewhere" into "the answer is one tap away," which is a meaningfully different proposition for civic engagement. The explicit design choice to avoid subjective scoring adjustments also speaks to a growing awareness that AI-assisted tools built on public data need reproducibility and auditability to be trusted, especially in politically sensitive domains where perceived bias could delegitimize the entire effort.
More broadly, this fits into a trend Anthropic and other AI labs have been keen to spotlight: individuals without formal technical backgrounds using conversational coding assistants to ship real, live software rather than toy demos. It echoes a wider "vibe coding" phenomenon where non-developers describe what they want in natural language and iterate with an AI on implementation, testing, and debugging. The Influence Registry project also demonstrates that this mode of development is increasingly being applied not just to prototypes or internal tools but to public-interest infrastructure — sites meant to be scrutinized, trusted, and used by strangers. As AI coding tools continue to mature, this kind of solo-builder civic tech may become more common, potentially reshaping who gets to build the tools that hold institutions accountable, while also raising new questions about verification, maintenance, and long-term reliability of software built this way.
Read original article →