From Idea to Dinner App in a Week: Micro Apps + LLMs

Build a deployable dining micro app in a week using LLMs, component tooling, and Rebecca Yu’s 7-day blueprint.

Stop wasting days debating dinner — ship a focused micro app in a week

Decision fatigue, fragmented group chats, and long research cycles are everyday pain points for developers and teams. Rebecca Yu’s seven-day dining app (Where2Eat) is a perfect blueprint: she used ChatGPT and Claude as copilots, built tiny, deployable components, and shipped a working product in a week. This guide turns her approach into a reproducible, developer-focused playbook for building micro apps with LLMs—fast, pragmatic, and production-minded.

Why micro apps + LLMs matter in 2026

By late 2025 and into 2026 the ecosystem reached a tipping point: LLM providers matured tool use and function-calling, edge and on-device inference became viable for low-latency experiences, and component tooling made it trivial to package UI as reusable units. That means you can now build a small, focused web app that: uses an LLM for decision logic, a tiny vector store or API for context, and a handful of well-tested components for UX.

Micro apps are intentionally narrow: they solve one problem for a small audience, iterate quickly, and have limited surface area for bugs and policy risk. For teams and solo devs aiming for an MVP or personal tooling, that’s gold.

Quick overview: 7-day blueprint

This is the inverted-pyramid plan: start with the minimum that delivers value, then iterate. The following schedule is tuned for developers who want production-ready output by day seven.

Day 0: Scope, data sources, and cost constraints
Day 1: UI skeleton and simple LLM prompt prototype
Day 2: Add context (RAG/vector DB) and refine prompts
Day 3: Build core components and Storybook stories
Day 4: Integrate LLM calls through serverless/edge functions for low latency
Day 5: Group features and realtime/shared state
Day 6: QA, accessibility, and cost/usage limits
Day 7: Deploy to edge, add observability, and ship

Day-by-day: Actionable steps, examples, and code

Day 0 — Define the MVP and constraints

Be ruthless. For a dining app like Where2Eat, the MVP scope could be: “Suggest three restaurants to a group based on preferences and recent chat consensus.” Capture these items:

User stories and acceptance criteria
Data sources: OpenStreetMap/Places API, optional review API, or a small CSV of favorites
LLM provider choice: ChatGPT for broad tool integrations, Claude for more controlled style, or a hybrid
Hosting constraints: edge functions for low latency, a vector DB for RAG (optional)
Budget: set a 7-day cost cap for LLM calls and hosting

Day 1 — Minimal UI + raw prompt prototype

Build a single page with a search box, a three-option chooser, and a minimal map or list. At the same time prototype prompt templates in the CLI or Postman so you can iterate quickly without the UI getting in the way.

Example system + user prompt template (start simple):

You are an assistant that recommends restaurants. Return a JSON array of three suggestions, each with name, short reason, and one-line rating rationale.

Serverless function stub (Node.js style, replace endpoints and keys):

export async function handler(req, res) {
  const userPrompt = req.body.prompt;
  const r = await fetch(process.env.LLM_ENDPOINT, {
    method: 'POST',
    headers: { 'Authorization': `Bearer ${process.env.LLM_KEY}`, 'Content-Type': 'application/json' },
    body: JSON.stringify({ model: 'gpt-4o', messages: [{ role: 'system', content: 'You are a restaurant recommender.' }, { role: 'user', content: userPrompt }] })
  });
  const json = await r.json();
  res.json(json);
}

Day 2 — Add context: small RAG and caching

LLMs are powerful but brittle without context. Use a tiny RAG layer to supply locality, favorites, or group history. In 2026, lightweight vector stores are common: Supabase Vector, Milvus, or hosted Pinecone/Weaviate. You can keep costs down with small embeddings and cached results.

Pattern: on user request, run a quick semantic search for context documents, then pass the top 3 snippets to the LLM as additional context. Cache embeddings and search results for identical queries.

Day 3 — Component-first UI: reusability and Storybook

Build small, focused components you can reuse or ship as packages: SearchInput, OptionCard, MapEmbed, and ShareInvite. Use Radix UI + Tailwind (or your preferred stack) for accessibility and speed. Add Storybook so every component has a living example — crucial for rapid iterations and handing off to designers.

Packaging options:

Publish as a NPM package if you want reuse across projects
Build as Web Components (Custom Elements) for framework-agnostic distribution
Offer an embeddable snippet that mounts an iframe for isolation

Day 4 — Integrate function-calling and tool access

By 2026 function-calling is standard. Let the LLM respond with structured calls (for example: callPlacesAPI, generateSummary) so you can keep business logic out of the natural language response. This improves reliability and makes it easier to test.

Example flow:

LLM returns a function call: { name: "search_places", arguments: { query: "sushi near me", filters: { price: 2 } } }
Your server executes the function against a real API or cached data
Your server returns the API result to the LLM for final phrasing if needed

Day 5 — Group features and shared state

Where2Eat’s core benefit is group coordination. Implement lightweight group state with Supabase or Firebase: a session URL, pinned preferences, and a simple voting mechanism. For realtime updates, use the platform’s realtime channels or server-sent events. Keep state normalized and small.

UX shortcut: avoid free-form input for group votes. Use quick actions like thumbs-up, thumbs-down, or rank-3 choices to limit LLM variability and speed decisions.

Day 6 — QA, accessibility, and cost controls

Run Playwright tests for flows, Vitest for unit logic, and Storybook for components. Audit prompts for hallucinations and edge cases. Add budget guards: hard caps on LLM calls per session, caching of repeated prompts, and sampled logs for manual review (redact PII).

Day 7 — Deploy and observe

Deploy the UI to a CDN-backed host (Vercel, Netlify, Cloudflare Pages). Deploy serverless/edge functions to Cloudflare Workers, Deno Deploy, or Vercel Edge Functions for low latency. Add basic observability: Sentry for errors, PostHog or Plausible for usage, and a simple Prometheus or hosted metrics for LLM call counts.

Prompt engineering patterns that speed iteration

Rebecca Yu’s “vibe-coding” relies on quick prompt loops. Use these patterns:

Constrained output: always ask for JSON or structured output to make parsing deterministic.
Few-shot templates: give 2–3 examples of desired output formats.
Tool calls: prefer function calls for deterministic operations (search, map lookup, scoring).
Safety layer: prepend a short guardrail that blocks policy-sensitive responses.

Example structured prompt for restaurant suggestions:

System: You are a restaurant recommendation engine. Provide exactly 3 suggestions as a JSON array. Each suggestion should have: id, name, short_reason, distance_meters, price_level (1-4).

Component architecture: shipping deployable pieces

Think in components that can be tested, versioned, and deployed independently. Two practical packaging patterns for micro apps:

Web Components: Build core widgets as custom elements. Consumers can drop a script and an element tag into any page.
Micro frontends: Use Module Federation or iframes for isolation and independent deployability. Good for teams that want separate release cadence.

Make your components production-ready by adding:

Storybook stories and visual regression tests
Small API surface and clear props/state contract
Types (TypeScript) and minimal CSS variables for theming

Two practical packaging patterns to explore are component registries and a component marketplace for deployable UI building blocks.

UX shortcuts that make the app feel polished fast

Defaults and progressive disclosure: Show three curated options first; let users dive deeper if they want.
Optimistic UI: Show provisional results while the LLM finalizes phrasing.
Choice architecture: limit options to reduce paralysis—3 choices beats 10.
Prefill with context: extract preferences from group chat messages with a short extraction prompt, but ask for confirmation.

Cost, privacy, and safety—practical guardrails

Even in a small app it’s critical to manage cost and risk.

Budget: set daily caps per user and global caps; batch and cache embeddings.
Privacy: never log raw conversation content unless explicitly allowed; redact PII before storing.
Security: sign and validate all server-to-server calls; limit CORS and require short-lived session tokens.
LLM safety: run a lightweight classifier on outputs if the app could cause harm (medical/legal topics).

Testing and observability for small apps

Test quickly and continuously. Each micro app should have:

Playwright end-to-end tests for the main happy path
Unit tests for prompt templates and parsing logic (Vitest/Jest)
Visual snapshots for main components (Chromatic or Percy)
Error monitoring (Sentry) and LLM-usage metrics with alerting

When to stop, iterate, or scale

Micro apps are transient by design. Use these rules to decide the next action:

If active users and engagement metrics grow, iterate and harden (add auth, billing, analytics).
If the app is used internally and reliable, package components as shared libraries to reuse across teams.
If the app becomes critical to users, raise the SLA: redundancy, backups for data, and a security review.

Example: core serverless function that uses function-calling

Below is a compact, provider-agnostic example of handling a function call returned by an LLM. Replace environment variables and SDK calls with your provider of choice.

export async function handler(req, res) {
  const prompt = req.body.prompt;
  // 1) Call the LLM
  const llmResp = await fetch(process.env.LLM_ENDPOINT, { method: 'POST', headers: { 'Authorization': `Bearer ${process.env.LLM_KEY}` }, body: JSON.stringify({ prompt, function_call: true }) });
  const llmJson = await llmResp.json();

  // 2) If the model requests a function, perform it
  if (llmJson.function_call) {
    const fn = llmJson.function_call.name;
    const args = llmJson.function_call.arguments;
    if (fn === 'search_places') {
      const apiRes = await fetch(`https://places.api/?q=${encodeURIComponent(args.query)}`);
      const results = await apiRes.json();
      // 3) Send results back to the model for final output or return to client
      return res.json({ suggestions: results.slice(0,3) });
    }
  }
  return res.json(llmJson);
}

Real-world tips from Rebecca’s workflow (applied)

Use the LLM as a pair-programmer: generate component scaffolding, then immediately review and refactor its output.
Favor short, iterative pushes. Ship a feature, watch usage for 48 hours, then refine.
Keep onboarding friction low: a join link plus a single click to share with friends beats complicated signups.

Future predictions (near-term, 2026–2027)

Expect these trends to shape micro app development in the next 12–24 months:

More on-device LLMs: low-latency personal micro apps running private models for privacy-sensitive features.
Component marketplaces: curated deployable widgets (LLM-enabled) you can plug into micro apps.
Standardized function schemas: industry-wide conventions for function-call contracts to make integration predictable.

Actionable takeaways

Start with a one-sentence problem statement and a 3-item acceptance criteria list.
Use a tiny RAG layer for context and rely on function-calling for deterministic operations.
Ship small, test early, and package components so they can be reused or embedded elsewhere.
Set cost and safety guardrails from day one.

Final thoughts and call-to-action

Rebecca Yu’s seven-day Where2Eat project proves the power of focused scope, LLMs as copilots, and component-driven development. With the patterns above you can reliably build a usable, deployable micro app in a week—production-ready and iteratable. Try the 7-day blueprint this week: pick a single decision problem (dinner, playlist, meeting time), scope the MVP, and follow the day-by-day plan. If you want a starter kit, examples, or a vetted component pack that includes a restaurant-picker widget and Storybook stories, sign up at javascripts.store or check our component marketplace for LLM-integrated UI building blocks. Ship faster, iterate smarter, and keep the app small.

From Idea to Dinner App in a Week: A Developer's Guide to Building Micro Apps with LLMs

Stop wasting days debating dinner — ship a focused micro app in a week

Why micro apps + LLMs matter in 2026

Quick overview: 7-day blueprint