Public-interest projects built with Claude

Nathanael Johnson · github.com/nathanaelhub · October–November 2026

Three tools for people who deal with public systems: families checking what help they might get, residents following their city council, and nonprofit staff starting to use AI. They're built on one idea: Claude handles the language, and code checks the facts.

Every number has a source. Figures are quoted from official pages, and tests fail if the source stops saying it.
Code checks Claude. Claude writes and explains; plain code does the math and checks the facts.
Measured, not assumed. Each project has an evaluation against known answers.
Says what it doesn't know. Unsettled rules, untested claims and limits are on the page.

The projects

Tennessee benefits screener

For families asking "might we qualify for SNAP, WIC or TennCare?"

Someone describes their household in plain words, in English or Spanish. Claude asks the questions it needs and explains the result. The eligibility math is plain code that Claude calls through MCP, with every threshold quoted from an official source.

98% vs 81%program calls right: 155/158 with the screener vs 110/135 with web search, 40 simulated households

19/22 vs 3/22SNAP amounts estimated within $5

$0.19 vs $0.49model cost per conversation

What building it found: Tennessee's SNAP income limit is unsettled. News reports say the state moved to 200% of poverty; the USDA's chart of states with that rule doesn't list Tennessee. The screener reports both, and a test fails the day the USDA adds Tennessee.
Limits: A screener, not a decision: results are "likely", "possibly" or "unlikely", and it always says to apply anyway. The evaluation uses simulated people, not real families.

Try it Code and evaluation

Nashville Council Digest

For residents asking "what did the Metro Council do, and what's in my district?"

Every bill and resolution in plain language, with its outcome and council district. Claude writes each summary; code checks every amount, number and name against the official text; a second Claude reading checks the meaning. If either check fails, the site shows the official title instead.

1,125bills and resolutions from 21 meetings, January–October 2026

6 real errorscaught by the second reading: it flagged 14 of 1,124 summaries the fact check had passed

0 vs 56 of 59planted name swaps caught by the code check vs the Claude reading

What building it found: The official record has its own errors: four, including an analysis naming the wrong nonprofit for a grant. And the city's "passed" flag reads "Fail" for a resolution the council adopted, so outcomes come from the recorded action instead.
Limits: The review of flagged summaries was done by Claude in the building session, not by an independent person; every case is in the repo to check. Districts are found for 180 of 194 site-specific zoning bills.

Read the digests Code and evaluation

Using Claude at a Nonprofit: a one-hour workshop

For nonprofit staff who don't write code, and the person who trains them.

A ready-to-run kit that teaches four habits: ask well, check before using, protect clients, and know when not to use it. It has a facilitator guide, slides, four exercises with answer keys, a privacy sheet quoting Anthropic's policies, and a quiz to give before and after.

Grade 2.6–5.8reading level of every attendee page, measured in code

$0.10to re-run every exercise through Claude before a workshop and score the answers

What building it found: I expected Claude to invent numbers when asked vaguely for a grant paragraph. It didn't: it left blanks like "[X] households" and asked for the real figures. The lesson moved to the person filling in the blanks, and the answer keys say so.
Limits: Not yet run with real staff. A one-page host flyer is out to find a first group.

See the kit Code and RUNBOOK

Checking the checker

An AI grader is only useful if you know how often it's wrong. Before relying on one, I measured it:

judge-audit grades answers whose right verdict is known in advance. When a Claude judge was wrong on all five votes, the item was wrong, not the judge: errors locked to an item meant three votes were worth 1.2 independent votes.
open-finance-mcp, a server that gives Claude free, citable SEC and Federal Reserve data, is where that question first came up.

How each project is handed off

Each repo has a README that says what was measured and what wasn't, tests that run on every change, and a runbook for the yearly updates: SNAP figures every October 1, new council meetings, changed policy pages. They're written so someone else can keep them running.