Journal · Building with Claude

A small team, building like a studio.

Between 11 and 28 October 2026, the Foundation’s small team made 2,410 commits across the eight codebases behind its apps, its platforms and this website, and one of those apps went from a blank repository to the App Store in four days. Most of that code was written with Claude Code. This is how, and the rules that make it trustworthy.

The Foundation’s approach rests on a simple bet: AI now lets a small team do what used to take a funded company. That only helps the people we serve if the result is reliable. A charity cannot afford a game that crashes, a form that loses a message, or a page that says something untrue. So the story here is less about how fast Claude writes code, and more about the system that lets us trust what it writes.

The numbers, counted from the code

Commits across eight repositories, 11 to 28 October
2,410
Pull requests reviewed and merged
369
Automated tests, Swift and TypeScript
4,800+
Days from Bonluma’s first commit to the App Store
4

The eight are the apps Charlie, Bonluma, Pawprint, Admit Path and Golden Hint, the Admit Path and Golden Hint web platforms, and this site. At least 1,396 of the commits carry a “Co-Authored-By: Claude” line. That is a floor, not the total: some of our repositories ask for no such line by policy. Bonluma, the match-three game now on the App Store, began as a product specification on the evening of 22 October and was released on 26 October, with 262 commits and 844 Swift tests behind it. Charlie, the largest, has 672 commits and nearly 1,500 tests across its Swift and TypeScript code.

The rules, written down where Claude reads them

Our repositories carry standing instructions that Claude Code reads at the start of every session, usually a short file called CLAUDE.md (Bonluma’s are in its worker guide). Ours are not style guides. They are the rules that protect the people who use what we build. A few, quoted directly:

  • “Every route has a test.” On this website, every address the server answers, and every email it sends, is checked by a test, and the tests never send real email.
  • Check before you finish. npm run check runs strict type checking, the full test suite inside the same runtime as production, and a dry-run build. The Admit Path and Golden Hint apps must build with no warnings and pass swift test before a commit.
  • Never invent a fact. When the site needs a fact nobody has confirmed, such as a release date or a figure, it shows a visible [TBC: …] marker and lists it for a person to fill in. A model that guesses plausibly is the risk, so the rule makes guessing impossible to hide.
  • Agents don’t touch production on their own. On this website: “Never deploy, run remote D1 commands or touch the Cloudflare dashboard from an agent session unless asked.”
  • Keep the README current in the same commit, including a dated decision log, so that a fresh session, or a new volunteer, can pick up the work from the repository alone.

The point of the rules is that nobody has to trust the model. They trust the tests, the reviews and the history, all kept in the repositories.

Many hands, one codebase

Claude Code can run as several agents at once. We give each one its own git worktree and branch, so parallel work never collides: one agent builds the game engine while another builds the accessibility layer. Bonluma’s worker guide sets the terms. Each worker edits only the files it owns, may commit but never push, and reports “build/test results as actually run”. Twenty-four of Bonluma’s merges came from those parallel branches.

Long tasks outlive any single session, so agents keep checkpoint files: what was done, the exact branch and commit, what failed, and the next steps. A new session reads the checkpoint, checks it against the real history and carries on.

Nineteen reviews of one website

After this site’s three-program rewrite, Claude reviewed it page by page, again and again, each time with fresh eyes, looking for anything untrue, unclear or stronger than the evidence. The commit history shows the rounds, from “Second review” to “Eighteenth review” on 24 October: a partner’s commitments marked as unconfirmed until they are in writing, a date corrected, a heading toned down. The nineteenth review found nothing to change.

Some of what those reviews caught shapes this very page. They are why the site says plainly that Anthropic is not a sponsor, why it never says Claude is donated, and why it doesn’t claim that Google Cloud runs our infrastructure when it doesn’t.

Claude as a user: Golden Hint’s connector

Claude also appears on the other side of our software. Golden Hint, the Foundation’s charity skills exchange, runs an MCP server at mcp.goldenhint.org, so a charity can use Golden Hint from inside Claude: browse listings, propose an exchange, read and answer its messages, or ask the community forum a question.

  • 48 tools in three access levels, public, read and write, each marked read-only or destructive so the assistant knows which actions need care.
  • Sign-in with OAuth, written for claude.ai custom connectors, or a personal token. Each person chooses read-only or read-and-write.
  • Safe by default. The server never logs tokens, and posting needs the organization to have been verified by Golden Hint staff first.
claude mcp add --transport http goldenhint https://mcp.goldenhint.org/mcp

Where Claude isn’t

Being specific means saying where Claude is not used, too. None of our apps calls Claude while you use it. The small amount of AI that runs inside them, such as Golden Hint’s “Help me write” or Pawprint’s daily story, uses Apple’s on-device models, so nothing leaves the phone. Claude is how we build; it isn’t in your pocket.

And people stay responsible. Decisions about facts, partners and legal text are recorded for a person to make, on this website agents can’t touch production unless asked, and anything a review couldn’t settle stays visibly marked on the page until someone does.

Why this matters for a foundation

None of this is a boast about speed. It is what makes our programs possible at all. Charlie, Bonluma and Pawprint can be free, with no ads and nothing to buy, because building them no longer costs what a studio costs. Every hour the code doesn’t need goes to the families, students and partners we work with.

Anthropic’s chief executive has written about AI helping the world make decades of progress in a fraction of the time. At the scale of one small foundation in New York, this is what that looks like in practice.

Built with

  • Claude Code: writing, testing and reviewing the code of every app and platform in this story.
  • Claude: the model behind it, and the assistant that connects to Golden Hint through its MCP server.

We build with Claude, made by Anthropic. Anthropic is not a sponsor of the Foundation, and nothing here suggests that it endorses the Foundation or its views. This story describes our own work; Anthropic did not write or approve it.

Updated : the rules now say which repositories each one applies to (the “rule one”, the standing-instruction files, the no-warnings rule and the production rule), and the heading counts nineteen reviews.

Build with us

Tools that let a small team build like a studio.

If your company makes tools for software teams and wants to see them used for public good, in the open and with the numbers published, we would like to hear from you.