I built 50+ projects in 97 days
using Claude Code and a terminal
Most nonprofits run on spreadsheets and duct tape. Not because they want to. Because custom software usually starts around $50,000.
So I ran an experiment: what happens if one person uses AI like a software team?
The setup
I'm Andrew. I run Frontier Commons, a nonprofit building tools for missions and international student ministry. Until January 2026, I had never shipped a website.
Then my friend Henry introduced me to Claude Code — Anthropic's AI that writes and runs code from your terminal. He sat next to me for the first few sessions. Not teaching syntax. Teaching how to think in prompts.
Within a week I'd shipped three tools my team had been requesting for months.
97 days later
50+ projects shipped. Not demos. The therapist directory gets 200+ searches a month.
A therapist directory. A benefits eligibility quiz. A 95K-article search engine. A Mario game that teaches cross-cultural skills.
What Claude Code actually does
You open a terminal. You type what you want. Claude writes the code, creates the files, runs commands, and fixes errors. You steer. It builds.
It's like dictating to a fast junior dev. You still make every decision — it just types faster than you.
My workflow
- Describe the problem — in plain English. "Build a searchable directory of 553 therapists with filters for language, specialty, and state."
- Claude builds the first version — usually 5–15 minutes.
- Improve it — "Make cards expandable." "Add a map." "Fix the mobile layout." Each instruction takes 30–90 seconds.
- Deploy — one command:
npx wrangler pages deploy .and it's live.
Most projects took 1–3 hours.
What I built
Tools nonprofits actually need:
- CAMH Directory — 553 therapists searchable by language, specialty, and state
- Missions Search — 95K articles across 28 sources
- Care Navigator — answer 8 questions, see benefits you qualify for
- IFI Dashboard — weekly nonprofit team dashboard
- Uncover — photo conversation tool for small groups
- CLI Dojo — RPG-style terminal training game
What surprised me
I wasted more time on vague prompts than on any bug. The projects that shipped fastest were the ones where I could describe the problem in one sentence.
- It gets faster. After ~10 projects, patterns appear. Projects 30–42 shipped 3× faster than projects 1–10.
- Iteration becomes free. Want a new design? 2 minutes. So you test 5 ideas instead of debating one.
- Single-file HTML handles more than you'd think. 15 of my 50 projects are one .html file.
- Deployment is trivial. Cloudflare Pages:
npx wrangler pages deploy .— 10 seconds. Figuring out what your team actually needs takes longer.
10 lessons after 50+ projects
- Solve a real problem. "My team needs a dashboard" beats "I want to learn React."
- Be specific in prompts. 'Build a searchable table' loses to 'Build a sortable table of 553 rows with filters for state, language, and specialty.'
- Ship ugly first. My best projects started as things I was embarrassed to share.
- Describe the problem, not the solution. Say 'users can't find therapists by language' instead of 'add a dropdown filter.'
- Use single-file HTML. Less complexity. Faster builds.
- Deploy early. Working links change conversations.
- Use a CLAUDE.md file. Claude reads it every session. Huge productivity boost.
- Learn the workflow, not the framework. Describe → build → review → refine → deploy.
- Pair with someone early. Saved me days of confusion.
- Build 10 things, not one perfect thing. Patterns appear around project 10.
Want to learn this workflow?
I turned everything above into a free training course. 10 levels. Start by installing your first terminal tools, end by shipping a full-stack app. 10 levels. You install tools, write prompts, and deploy a real site by level 7.
CLI Dojo — RPG-style terminal training. XP, boss challenges, and a curriculum that takes you from zero to deployed.
Must-have skills
Claude Code has an official skills library — drop-in modules that give Claude specialized abilities. These are the ones I use constantly:
- Frontend Design — Forces real design decisions — typeface pairing, color system, spacing — instead of the default blue-gradient look. Forces bold aesthetic choices: editorial, brutalist, retro-futuristic. Picks real typefaces, builds cohesive color systems, adds purposeful animation. The single biggest upgrade to how my sites look.
- Skill Creator — You describe what you want. It drafts the skill, tests it, and iterates until the output matches. I used it to build my
/roastand/deploycommands. - Claude API — Detects your project language, recommends the right API surface (single call, workflow, or agent), and writes the integration code. Essential for adding AI features to any app.
- MCP Builder — Guides you through building MCP servers — the protocol that lets Claude talk to external services. Four-phase workflow: research, implement, test, evaluate.
Install any of them by cloning the skills repo and symlinking into ~/.claude/skills/. Five minutes to set up. I use them every session.
My stack
- Frontend — HTML, CSS, vanilla JS. React only when necessary.
- Hosting — Cloudflare Pages. One command to deploy.
- AI — Claude Code for all development.
- Data — Python scripts + Cloudflare D1.
- No Docker. No CI/CD. Add complexity only when needed.
~/Documents/Claude/ ├── camh-directory/ ├── care-navigator/ ├── claude-portfolio/ ├── crossings-mario/ ├── ifi-dashboard/ ├── missions-search/ ├── uncover/ └── ... 50+ more
Each project gets its own folder.
The real productivity hack
Every project includes a CLAUDE.md file. Claude reads it at the start of every session. 10 seconds and Claude knows the whole project.
# CAMH Directory ## Stack - Static HTML + vanilla JS ## Deploy npx wrangler pages deploy . --project-name=camh-directory ## Current State - 553 counselors, 40 states, 26 languages
Claude instantly understands the project.
I also built slash commands — custom skills Claude runs on demand:
/deploy— commit + deploy/roast— 10 expert personas critique the project/resume— Claude finds the next task/wrapup— commit, docs, summary email
After enough projects, starting a new one takes 5 minutes.
The full toolkit
Claude Code ships with 93 built-in commands. Here are the ones that changed how I work:
/resume Pick up where you left off/compact Compress context when it gets long/rewind Undo conversation and code changes/diff Review all changes before committing/plan Read-only planning — no file edits/fast Same model, 2.5× output speed/agents Spawn sub-agents for parallel work/review Code review a PR/security-review Security audit of uncommitted changes/simplify 3-agent architecture + perf review/hooks Run scripts on lifecycle events/mcp Connect to external services/loop Run a prompt on a recurring schedule& task Background a task and keep workingPlus my custom commands — /deploy, /roast, /wrapup, /next — built with the skills system. 93 commands out of the box. Unlimited custom ones on top.
The Complete Field Manual — every Claude Code command, shortcut, and workflow in one reference page.
The bigger idea
A nonprofit with one person and a terminal can now build tools that used to require a contractor and $50K.
The hard part is still the same: knowing what your people actually need. AI just removed the excuse that you can't build it.
If your nonprofit has a spreadsheet it wishes were a tool, email me.