Let the AI Set Things Up
Last reviewed
BeginnerWhat you'll learn
~15 min- Understand what 'scaffolding' means and why it used to be the hardest part of development
- Write a clear scaffolding prompt that describes what you want (not how to build it)
- Verify that an AI-generated project has the right structure and key files
- Apply the vision-then-verification workflow to any project type
- Create a short project memory file (AGENTS.md or CLAUDE.md) right after scaffolding
By the end of this lesson, you’ll be able to describe a project to an AI CLI tool and get back a complete, working foundation — then verify that it built what you asked for.
The Mental Model: Hiring a General Contractor
Imagine you want to renovate your kitchen. You do not need to know how to run electrical wiring, install plumbing, or frame walls. You hire a general contractor, describe what you want (“an open layout with an island, gas range, and pantry”), and they handle the implementation.
AI scaffolding works the same way. You are the homeowner with the vision. The AI is the contractor with the skills. Your job is to clearly describe what you want and then inspect the work when it is done.
What makes a good homeowner (orchestrator):
- Clear about what they want (“I need a kitchen island with seating for four”)
- Inspects the work (“Are the outlets where I asked for them?”)
- Gives feedback (“Move the pantry to the other wall”)
What they do NOT need to know:
- How to read a wiring diagram
- Building code specifications
- Plumbing pipe sizing calculations
The Liberating Truth
Here is a sentence that would have been absurd five years ago and is completely true today:
You can get started without mastering setup details. Over time, learning some setup basics will help you debug and customize faster.
Setting up a new project used to be the hardest part of development. Choose a framework, configure the build tools, set up testing, configure linting, create the folder structure, install dependencies, wire up the database… Each step required specialized knowledge and could take hours of debugging.
Now? You describe what you want, and the AI does all of it.
Scaffolding = Letting the AI Build the Foundation
“Scaffolding” means creating the initial structure of a project — the folders, configuration files, starter code, and dependencies. It is the construction equivalent of pouring the foundation and framing the walls.
Let’s walk through what scaffolding looks like from start to finish with an AI agent (such as Claude Code, Antigravity, or your preferred tool).
Step 1: Say what you want. Open your tool in the folder that will hold the project (the AI creates the project folder inside it). Switch it to plan mode first, Shift+Tab until the status bar shows plan (Module 6’s plan mode section: it proposes before it builds). Then send a short goal on its own turn (a turn is one message you send). Describe the outcome, not the tools:
I want a simple website for a university research lab: a homepage, a team page, a publications page, and a contact page. Use placeholder content I can easily replace, and ask me what you need before you plan.The agent will ask questions, such as the look you want, how many team members, and whether the publications need search. Answer in plain English. Approve the plan once it lists the four pages. Setup that can take a developer an hour by hand usually takes a few minutes this way (your timing will vary with the project and the tool).
Step 2: Ask the agent to show you what it built. After the agent finishes, ask:
List the pages you created and tell me which files matter most.
You should see four pages that match your request (a homepage, team, publications, and contact), plus a README.md and a src/ folder.
Step 3: Preview it and check. Ask:
Start a local preview and tell me the address to open.
If a tool is missing, the agent will offer to install it. Approve that step. Open the address it gives you and click through every page. Does the navigation work? Are there four pages? Does it look the way you described?
Step 4: Give feedback. If something is not right, say so in plain words:
The team page is showing team members in a list. I want a grid of photo cards instead, with a name, role, and one-line research interest on each.This is the build loop at its smallest: plan, approve, build, check, keep. Think Before You Prompt turns this into a full loop: the agent asks questions first, and the project’s checks decide when a step is done.
The Key Insight
You do not need to understand:
- How the styling engine is wired into the build
- What the framework configuration file does internally
- How compilation or bundling is configured
You need to understand:
- What you want to build (a website with specific features)
- What it should look like (dark theme, clean design)
- What technologies to use (let the AI choose, or name a specific stack if you have one)
If you do not have a preference, it is totally fine to say “Build me a personal portfolio website. Choose whatever tech stack makes sense.” The AI will pick reasonable defaults. As you gain experience, you’ll develop opinions about tools and frameworks — then you can be more specific.
When you already have a tech stack in mind
Once you develop preferences (or when your team already uses a specific framework), you can name those technologies directly in your plan-mode prompt and approve the plan it proposes:
Create a new web application with these specs:- Astro framework with React components- Tailwind CSS for styling- Dark theme- A homepage with a hero section- A navigation bar with Home, About, and Contact links- TypeScript for type safetyThe AI will:
- Create the project folder
- Initialize
package.json - Install all dependencies
- Create the Astro configuration
- Set up Tailwind with the dark theme
- Create the layout with navigation
- Create the homepage with hero section
- Set up TypeScript configuration
All from one prompt.
🧬In Your Field: Biotechclick to expand
Plain-English goals for bioinformatics projects (after switching to Plan mode):
Lab inventory tracker:
Help me build a browser-only lab reagent tracker that flags items expiring within 30 days and exports to CSV. Ask me what columns and checks we need before writing code.Sequence analysis dashboard:
Help me build a clean scientific dashboard where I can upload or paste a FASTA sequence to inspect length, GC content, and base composition, including one sample sequence. Ask me what you need first.Describe what you need in domain terms, then check the agent’s proposed stack and validation steps in its plan before approving.
🏛️In Your Field: Government Devclick to expand
Plain-English goals for public-sector projects (after switching to Plan mode):
Public services directory:
Help me plan a mobile-friendly searchable directory for public services (permits, licenses, records, benefits) targeting WCAG 2.2 AA accessibility, seeded with 10 mock entries. Ask me anything unclear before building.Internal status dashboard:
Help me plan an internal project status dashboard with department and status filters, summary counts, and 15 mock projects across four departments. Ask me what details you need first.Notice the accessibility goal in the first starter: for public-facing government tools, state WCAG 2.2 AA up front and verify keyboard and contrast checks in the plan.
🎯In Your Field: All Tracksclick to expand
Beyond web apps: command-line tools
You can scaffold non-web tools the same way in Plan mode:
Help me plan a command-line CSV analyzer that reads a file path, prints row and missing-value stats, and saves a PNG bar chart of top values in a chosen column. Ask me what you need first.Let the agent propose the libraries and folder layout in its plan, then approve once it looks clean.
Real Examples
Example 1: Blog
Prompt:
Create a blog using Astro with:- Markdown content in a content/ folder- Homepage showing recent posts- Individual post pages- Dark theme with good typography- Tags on postsWhat the AI creates:
my-blog/├── src/│ ├── content/posts/│ │ ├── first-post.md│ │ └── second-post.md│ ├── layouts/│ │ └── BlogLayout.astro│ ├── pages/│ │ ├── index.astro│ │ └── posts/[slug].astro│ └── styles/global.css├── astro.config.mjs└── package.jsonA complete blog. Ready to customize. (Your exact file structure may differ slightly depending on the AI tool and model — verify the required pages and features are present rather than expecting identical filenames.)
Example 2: Dashboard
Prompt:
Create a data dashboard with:- React + Vite- A sidebar navigation- Three pages: Overview, Analytics, Settings- Charts using a charting library- Mock data for demonstration- Dark themeExample 3: API Server
Prompt:
Create a REST API with:- Node.js and Express- Endpoints for users (CRUD)- SQLite database- Input validation- Error handling- Basic test suiteIn each case, the AI builds the complete scaffolding and initial functionality. Your job is to refine, customize, and extend.
The Orchestrator’s Advantage
Here is why this matters so much:
Without AI scaffolding:
- “I want to build X” —> days or weeks of learning before you can start —> many people quit
With AI scaffolding:
- “I want to build X” —> AI creates a working scaffold in minutes —> you start customizing immediately —> you learn by doing
This flips the entire learning model. Instead of learning theory first and building later, you build first and learn the theory as needed. When you see the code the AI generates, you naturally start understanding patterns. When something breaks, you learn debugging. When you need a feature, you learn the relevant concept.
The AI just generated an entire project from your description. If that feels wrong — like you skipped the hard part — reconsider what “the hard part” actually is. Writing boilerplate config files isn’t the hard part. Knowing what to build, describing it clearly, and verifying the result is right? That’s the hard part. That’s orchestration. You didn’t skip anything — you did the part that matters.
Setting up project memory
Right after scaffolding a project is the best time to create a project memory file — a document the AI reads automatically at the start of every session.
Each tool drafts or reads its own file:
- Claude Code:
CLAUDE.md(type/initand it scans your freshly scaffolded project and drafts one) - Codex:
AGENTS.md(type/initto create one) - Copilot CLI:
.github/copilot-instructions.md(/initsets it up); it also readsAGENTS.md,CLAUDE.mdandGEMINI.md - Antigravity:
AGENTS.md(orGEMINI.md) in your project root; it has no/init, so ask it in plain words to draftAGENTS.mdfrom the project - Other tools: Check your tool’s documentation for the current convention. If your tool does not support automatic memory files, keep a
PROJECT_NOTES.mdand paste the relevant context at the start of each session.
Add your project’s purpose, naming conventions, and anything the AI should know across sessions. A few lines is enough to start. Tool-specific conventions can change between versions, so verify with your tool’s docs if something is not working.
Keep it short, and keep one file. Once the draft exists, send this on the next turn:
Review what you just wrote. Keep the build, test and lint commands and the rules this repo cannot tell you; cut anything a reader could learn from the code.Codex, Antigravity and Copilot CLI all read AGENTS.md, so keep that as the one shared file if you or your team use more than one tool. Claude Code reads CLAUDE.md first and falls back to AGENTS.md only when there is no CLAUDE.md, so if you have both, add the line @AGENTS.md to CLAUDE.md and Claude Code loads the shared file through it.
The structure is fresh, the conventions are established, and the AI can capture them accurately. Waiting until later means you have to reconstruct context that already existed.
When Things Go Wrong
When Things Go Wrong
Use the Symptom → Evidence → Request pattern: describe what you see, paste the error, then ask for a fix.
Run this on your machine
- Make a fresh, empty folder for a small project and open it in your agent (the Desktop App or the CLI).
- Switch to plan mode.
- Send one or two plain sentences about something you would like, for example: “I want a simple personal portfolio website with an about page and a projects page.”
- Answer the questions the agent asks. It will propose a plan; read it and approve it.
- Let it set up and run the project, then open the address it gives you in your browser and check the live result.
Notice what you did not have to do: pick a framework, list every file, or write a long specification. The agent made those choices and asked you about the rest.
Key Takeaways
- Scaffolding is letting the AI create the entire project foundation — folders, config files, starter code, dependencies — from a single prompt
- You are the homeowner, not the contractor: describe what you want, inspect the result, give feedback
- A good scaffolding prompt describes what (pages, features, design) not how (webpack config, CSS reset, TypeScript settings)
- Always verify after scaffolding: ask the agent to list what it built, start a preview, and click through every page
- The build loop is plan, approve, build, check, keep, repeat until it matches your vision
🔍What if the AI picks a tech stack you don't want?click to expand
Sometimes the AI will default to a stack you have not heard of or do not want. You can always be explicit: “Use Astro, not Next.js” or “Use plain JavaScript, not TypeScript.” If the AI already created the project with the wrong stack, you have two options: (1) ask it to convert the project (“Convert this from Next.js to Astro”), or (2) start fresh with a more specific prompt. Starting fresh is usually faster for major tech stack changes.
What's the most important thing you need to know to scaffold a project with an AI tool?
After the AI scaffolds your project, what should you do first?