llms.txt
Compare with other sites' llms.txt files →# Steamfield
> Steamfield is a small software studio. We build OpenData, OpenChart, StatGlass, Stacks, rr, Decks and foreman, and write about how they work.
## Products
- [OpenData](https://tryopendata.ai): Connect OpenData to Claude or ChatGPT and ask about schools, housing, health or jobs. It finds the right public datasets, does the math, and draws the chart inside the conversation.
- [OpenChart](https://tryopendata.github.io/openchart/): Describe a chart as JSON and OpenChart handles the scales, label placement, responsive layout and dark mode, then renders it in React, Vue, Svelte or plain JavaScript.
- [StatGlass](https://statglass.steamfield.io): StatGlass lives in the menu bar of Apple Silicon Macs, reads CPU, GPU, memory, power, network, disk and battery once a second, and keeps up to 30 days of it.
- [Stacks](https://stacks.steamfield.io): Stacks writes the starter files and a build brief for a new project, hands it to Claude Code, Codex or Cursor, and runs the project's own checks until they pass. This site was initially bootstrapped with the Stacks CLI.
- [rr](https://rr.steamfield.io): rr syncs your working tree to remote machines over SSH, runs the tests there, and streams the results back, split across as many hosts as you have.
- [Decks](https://decks.steamfield.io): Decks is a Slidev addon that lets a coding agent build a presentation from a brief. It ships the slide components and chart tools, and its init command installs the rules and skills that tell the agent how to use them.
- [foreman](https://foreman.steamfield.io): foreman lets Claude Code delegate coding work to Codex models. Claude writes a contract for each task, and foreman starts each delegate in its own git worktree.
## Blog
- [How to block dangerous commands in Claude Code](https://steamfield.io/blog/block-dangerous-commands-claude-code-hooks): Add deny rules to settings.json, then a PreToolUse hook on Bash that reads the command from stdin and denies destructive ones before they run.
- [How to make a presentation with Claude Code](https://steamfield.io/blog/build-slide-deck-with-coding-agent): Claude Code can write a whole Slidev deck in Markdown. How to keep it from producing bullet walls, generic AI prose and numbers nobody can trace.
- [How to connect Claude to government data](https://steamfield.io/blog/connect-claude-to-public-government-data): Add an MCP server to Claude and it can search public sources like Census, BLS and FRED, run SQL on the tables and link each number to its dataset.
- [How to design a CLI for AI agents](https://steamfield.io/blog/design-cli-for-ai-agents): Adding --json is the start. An agent-friendly CLI also needs stable event streams, parsed failures with file and line, and error codes that name the fix.
- [How to generate charts with an LLM](https://steamfield.io/blog/generate-charts-with-llm-json-spec): Have the model write a compact JSON chart spec, validate it against your real data, and send field-level errors back so it can fix the chart in one retry.
- [How to write a CLAUDE.md file the agent follows](https://steamfield.io/blog/how-to-write-claude-md-agents-md): Keep CLAUDE.md under 200 lines, list the exact commands, import AGENTS.md for other tools, and move rules that must always hold into hooks and settings.
- [Make a coding agent run tests before it's done](https://steamfield.io/blog/make-coding-agent-run-tests-before-done): Instructions only ask an agent to run tests. One verify command for every gate, plus a Stop hook that blocks the finish until it passes, makes it happen.
- [How to run multiple coding agents in parallel](https://steamfield.io/blog/run-parallel-coding-agents-without-collisions): Give each agent its own git worktree, port range and test database, merge through one integrator, and keep their test runs off the same machine at once.
- [How to run tests on a remote machine over SSH](https://steamfield.io/blog/run-tests-on-remote-machine-over-ssh): Copy your working tree with rsync, run the suite over ssh, and copy the reports back. The one-liner, where it breaks, and how to split a suite across hosts.
- [How to see CPU usage history on a Mac](https://steamfield.io/blog/see-mac-cpu-usage-history): Activity Monitor only shows live CPU data. Here is how to find what used your Mac's CPU earlier, log it from Terminal, or keep weeks of history.
- [How to use Claude Code and Codex together](https://steamfield.io/blog/use-claude-code-and-codex-together): We use Claude Code to plan and review, and Codex to code in worktrees. Here's the contract, review loop and token usage from 58 completed runs.
- [Drawing the charts on this site at build time](https://steamfield.io/blog/charts-at-build-time): Every chart on steamfield.io is an OpenChart spec rendered to static SVG at build time, so pages ship no charting code and figures work without JavaScript.
- [Why Steamfield](https://steamfield.io/blog/why-steamfield): Steamfield is a small studio that builds its own products and takes no client work. This post covers what they are and how each one grew out of another.
## Optional
- [RSS feed](https://steamfield.io/rss.xml): full text of recent posts
- [Privacy](https://steamfield.io/privacy): what the site measures and how to opt out