In short
Build a tool that helps people rewrite emails, SOPs or captions in the tone they want. It's a great first LLM project because it uses system prompts, templates, streaming and a simple UI together.
A tool that helps users draft, edit or rewrite text with an LLM — a practical project that exercises prompt design, streaming responses and a real UI.
Think of it like this
The helpful senior
Before you send that cover letter, you show it to a senior who's good with words. She doesn't write it for you. She tightens sentences, fixes tone and asks 'what did you actually mean here?' Your writing assistant should be that senior, not a ghost-writer.
Watch it happen
Build plan
1/5A simple page: text box, tone dropdown (formal, friendly, concise) and a Rewrite button.
The big ideas
What you need to know
Tone and length controls
Dropdowns for tone and length make prompts predictable and the tool easy to use.
Show a diff
Highlight what changed so users learn and can reject edits they don't like.
Keep meaning
Tell the model not to add facts, numbers or claims that weren't in the original.
Protect the API key
Call the model from your backend. Never put the key in frontend JavaScript.
See it in code
Walk through the code
1SYSTEM = ("You are a writing coach. Improve clarity and tone. "2 "Never add facts, numbers or claims not in the original text.")3 4TEMPLATES = {5 "rewrite": "Rewrite the text in a {tone} tone. Keep it under {words} words.",6 "shorten": "Shorten the text to about half its length. Keep the key points.",7 "grammar": "Fix grammar and spelling only. Do not change the wording otherwise.",8}9 10def build(action, text, tone="friendly", words=150):11 instruction = TEMPLATES[action].format(tone=tone, words=words)12 return [13 {"role": "system", "content": SYSTEM},14 {"role": "user", "content": f"{instruction}\n<text>{text}</text>"},15 ]1/3The system prompt sets the role and the most important rule: don't invent content.
The core of the app: one system prompt, one template per button, and a function that builds the messages.
Save yourself the pain
Common mistakes
Oops
Letting the assistant 'improve' text by adding achievements the user never had.
Do this instead
Forbid new facts in the system prompt and test with resumes and SOPs to make sure it holds.
Oops
Calling the model directly from the browser with your API key.
Do this instead
Route calls through a small backend endpoint, and add basic rate limiting.
Your mission
First milestone: one button works
Quick check
Why should your writing assistant's API key live on the server, not in the frontend?
Go deeper
- 1Build a small app with editing modes (shorten, formalize, fix grammar) as distinct prompts.
- 2Stream the model's response to the UI instead of waiting for the full output.
- 3Add a system prompt that keeps the assistant's tone consistent across modes.
More projects at this stop
Structured Data Extractor
Stuck on this topic? Ask an Insider
Get 1:1 guidance from people who've walked this exact path — on the InsideEdge app.
