Bolt.new — Module 22.6

How This Lesson Fits the Module & Volume

Lectures 1–5 were developer daily drivers: forks, plugins, terminal agents. Bolt.new (StackBlitz) starts the prompt-to-running-app cluster: a browser IDE that generates and runs full-stack JavaScript/Node in WebContainers—no local install. That is a different product from Vol. 21 coding assistants embedded in a pre-existing monorepo, though the HITL rule still holds: you own what you export.

Vol. 18 deployment / Docker remain the production path; Bolt is a prototype and teaching surface. Vol. 15 agents appear as the chat that writes files, not as your production orchestrator. Next: Lovable (similar “describe an app” job, different stack gravity).

Learning Objectives

By the end of this lesson, students should be able to:

Define Bolt.new as a browser prompt-to-app environment on StackBlitz WebContainers.
Contrast it with agentic desktop IDEs (Cursor) and autocomplete plugins (Copilot).
State WebContainer limits: in-browser Node, not arbitrary native binaries or huge monorepos.
Plan export → git → Vol. 18 deploy instead of treating the preview URL as production.
Apply privacy: do not paste customer PII or prod secrets into the Bolt prompt.
Know when to pick Bolt vs Lovable vs Replit vs a real IDE.

Definition

Bolt.new is StackBlitz’s AI app builder: you describe a web app in chat; the agent writes project files and runs them in the browser via WebContainers (a browser-native Node-like runtime). You get a live preview, can iterate in chat or the editor, and typically export to GitHub or a host. Bolt is closer to an agentic scaffold + preview than to ghost-text autocomplete. SKUs (tokens/credits, Pro) evolve—read the current plan page; do not invent dollar prices.

Not Your Daily Driver IDE

	Bolt.new	Cursor / Copilot
Starting point	Empty or generated app from a prompt	Existing repo you already clone
Runtime	WebContainer in the browser	Local / remote dev machines
Autocomplete vs agent	Agent-first (chat writes many files)	Tab + optional Agent
Best job	MVP, demo, workshop, spike	Ongoing product engineering
Vol. 18 ship	Export then Docker/CI	Already on git + CI

WebContainers buy

Zero install for students
Instant preview
Shareable spike URLs

WebContainers cost

No arbitrary native addons
Weak fit for huge polyglot monorepos
Preview ≠ hardened prod

HITL still

Read the generated auth code
Do not commit secrets
Run your own tests after export

Privacy and “Repo” Context

There is often no prior private monorepo—the prompt is the spec. That makes prompt hygiene the privacy control: never paste customer dumps, API keys, or internal architecture you could not publish. If you import a GitHub repo into Bolt, treat it like any cloud IDE: assume files may be sent to the model vendor; use public or dummy data. Vol. 20 privacy does not pause because the UX is fun.

Catalog Snapshot (Qualitative)

Dimension	Bolt.new	Lovable	Replit AI
Pricing posture	Subscription + generation/token or credit meters—confirm live sheet	Credits / generations + seats	Workspace plan + Agent credits/usage
Strengths	WebContainer preview; StackBlitz DNA; fast JS/TS full-stack spikes; export to git	Opinionated full-stack + hosted backend gravity (often Supabase-class)	Cloud VM-like workspace, collab, in-place deploy
Weaknesses	Browser runtime limits; not a JetBrains daily driver; easy to confuse preview with prod	Stack conventions / lock-in risk; weaker as “open any repo” IDE	Cloud gravity; residency questions; not local-offline first
Agentic vs autocomplete	Agent-first scaffold	Agent-first scaffold	Agent + in-editor assistant (Ghostwriter-class history)
Privacy / context	Prompt + imported files to cloud model; dummy data only	Same cloud-builder caution	Cloud workspace policy + org controls

Prompt → Export Checklist

# bolt_brief.md — educational prompt pattern (no PII, no real keys)
Build a small FastAPI-style *frontend* dashboard in Vite + React:
- Page: list mock "orders" from in-memory JSON (no real customers).
- Form: add an order (client-side only).
- Tests: at least one Vitest for the total calculator.
- Do not include .env secrets. Use placeholder NEXT_PUBLIC_API_URL.
- Accessible labels; responsive layout.

# after_export.py — Vol. 18 hand-off (systems you maintain)
CHECKS = [
    "git clone export && npm test",
    "scan for hardcoded keys",
    "Dockerfile or host config reviewed",
    "preview URL is not the production canonical",
    "HITL code review before merge",
]

def ready_to_deploy(results: dict) -> bool:
    return all(results.get(c, False) for c in CHECKS)

When to Pick Bolt.new

Pick Bolt when

Workshop / classroom with no local toolchain.
JS/TS web MVP that benefits from instant preview.
You will export to GitHub and graduate to CI.
StackBlitz/WebContainer constraints are acceptable.

Pick something else when

Existing polyglot monorepo → Cursor / Copilot / Claude Code.
You want hosted DB/auth opinionation → Lovable.
You need multiplayer cloud VM + many languages → Replit.
Native/mobile/GPU training → not Bolt.

Related Lectures

Lecture	Role
Coding assistants	HITL still applies to generated apps
Agentic workflow	Chat-writes-files is an agent pocket
Deployment / Docker	Where Bolt previews must graduate
Lovable / Replit AI	Sibling builders

Common Misconception

“Bolt replaces Cursor for our 400-service monorepo.” Wrong starting point. Second: the preview URL is production (TLS, scaling, secrets, audit). Third: WebContainers run any Docker image. Fourth: pasting prod logs into Bolt chat is harmless. Fifth: inventing credit prices. Sixth: generated auth is automatically Vol. 20 compliant.

Knowledge Check

Short Answer: What runtime makes Bolt.new able to run apps in the browser? Answer: StackBlitz WebContainers.
True/False: Bolt.new is primarily ghost-text autocomplete in JetBrains. Answer: False—it is an agent-first browser app builder.
Multiple Choice: Bolt is a best fit for: (a) JS/TS MVP spikes, (b) air-gapped Fortran HPC, (c) unsupervised kube prod. Answer: (a).
Short Answer: What should you do before calling a Bolt app “shipped”? Answer: Export to git, test, review, and deploy via Vol. 18 (not the preview URL alone).
True/False: You may paste real customer PII into the Bolt prompt for realism. Answer: False.
Multiple Choice: Pricing posture is: (a) subscription + generation/credit meters, (b) a fake $5 figure, (c) only self-host n8n. Answer: (a).
Short Answer: Name one WebContainer limitation. Answer: Arbitrary native binaries / huge polyglot monorepos / preview ≠ prod (any one).
True/False: Lovable and Replit are sibling prompt-to-app tools in this module. Answer: True.
Multiple Choice: Vol. 18 role after Bolt is: (a) real deployment/CI, (b) more ghost text, (c) BLEU. Answer: (a).
Short Answer: When pick Cursor instead of Bolt? Answer: Ongoing work on an existing large repo / local toolchain.

Key Takeaways

Bolt.new is StackBlitz AI + WebContainers: prompt-to-preview, not a monorepo daily IDE.
Agent-first scaffold; autocomplete IDEs remain for existing codebases.
Privacy = prompt hygiene; export then Vol. 18 deploy.
Compare qualitatively with Lovable and Replit on runtime and lock-in.
Next: Lovable.

Trainer’s Guide

Lab: Students generate a mock orders dashboard (no real PII). Export. Run tests locally or in CI. List three WebContainer limitations they hit.

Compare: Same prompt in Lovable next hour—note backend opinions vs Bolt’s in-browser runtime.

Recap: Bolt.new opens the browser-builder trio: instant WebContainer previews, agentic scaffold, Vol. 18 after export. Continue to Lovable.