.png)
4 months ago
18
A lightweight utility for building clean, composable, and maintainable LLM prompts using tagged template literals.
import { prompt } from 'llm-prompt-tag';
const problems = [
"String concatenation",
"Conditional logic mess",
"Formatting inconsistencies",
"Hard to maintain"
];
const solutions = [
"Tagged templates",
"Conditional sections",
"Automatic formatting",
"Composable structure"
];
const includeDebugInfo = false;
const msg = prompt()`
You are a helpful AI assistant. It can be really annoying to put all the conditional data in your prompts.
${prompt('Problems with creating prompts')`${problems}`}
${prompt('Solutions from llm-prompt-tag')`${solutions}`}
${prompt('Debug Info', includeDebugInfo)`This section only appears when debugging`}
Now I can make complex, conditional, formatted prompts I can still read thanks to llm-prompt-tag!
`;
console.log(msg);
Output:
You are a helpful AI assistant. It can be really annoying to put all the conditional data in your prompts.
==== Problems with creating prompts ====
String concatenation
Conditional logic mess
Formatting inconsistencies
Hard to maintain
==== End of Problems with creating prompts ====
==== Solutions from llm-prompt-tag ====
Tagged templates
Conditional sections
Automatic formatting
Composable structure
==== End of Solutions from llm-prompt-tag ====
Now I can make complex, conditional, formatted prompts I can still read thanks to llm-prompt-tag!
- ✅ Native JavaScript/TypeScript tagged template syntax
- ✅ Automatic whitespace and formatting cleanup
- ✅ Optional section headers
- ✅ Conditional rendering
- ✅ makePrompt generator for type-safe extensibility
- ✅ Smart array formatting with customizable separators
- ✅ No dependencies
npm install llm-prompt-tag
import { prompt } from 'llm-prompt-tag';
const result = prompt('Instructions')`
Write a short, direct note using the user's own words.
`;
console.log(result);
Output:
==== Instructions ====
Write a short, direct note using the user's own words.
==== End of Instructions ====
const showHelp = false;
const result = prompt('Help Section', showHelp)`
If you get stuck, refer to the user's original notes.
`;
console.log(result);
Output:
(empty string - no output because condition is false)
With condition true:
const showHelp = true;
const result = prompt('Help Section', showHelp)`
If you get stuck, refer to the user's original notes.
`;
console.log(result);
Output:
==== Help Section ====
If you get stuck, refer to the user's original notes.
==== End of Help Section ====
Some people may prefer to rename the function for two distinct purposes: a simpler tag p for the overall template, and a separate one for sections.
This approach can be more readable and semantic:
import { prompt } from 'llm-prompt-tag';
// Use 'section' for structured sections
const section = prompt;
// Use a plain wrapper for the main content
const p = prompt();
const someVariable = "User prefers dark mode";
const notes = "Remember to save work frequently";
const result = p`You are a helpful AI assistant.
${section('User Settings')`${someVariable}`}
${section('Notes')`The user's notes are: ${notes}`}
Whatever approach feels cleaner to you!`;
console.log(result);
Output:
You are a helpful AI assistant.
==== User Settings ====
User prefers dark mode
==== End of User Settings ====
==== Notes ====
The user's notes are: Remember to save work frequently
==== End of Notes ====
Whatever approach feels cleaner to you!
If you have custom objects you're putting into your prompts you can make that easier/cleaner by using the makePrompt generator with a list of types and their to-text formatters.
Example:
import { makePrompt } from 'llm-prompt-tag';
class Note {
constructor(public title: string, public content: string) {}
}
const note2Text = (note: Note) => `• ${note.title}\n${note.content}`;
const promptWithNotes = makePrompt({
Note: note2Text
});
const note = new Note("LLM Summary", "LLMs are transforming software development.");
const result = promptWithNotes('User\'s Notes')`
Here's a user note:
${note}
`;
console.log(result);
Output:
==== User's Notes ====
Here's a user note:
• LLM Summary
LLMs are transforming software development.
==== End of User's Notes ====
To keep templates simpler, arrays of registered types are automatically formatted. By default, items are separated with double newlines for clean readability:
class Note {
constructor(public title: string, public content: string) {}
}
const promptWithNotes = makePrompt({
Note: (note) => `• ${note.title}\n ${note.content}`
});
const notes = [
new Note("Meeting", "Discuss project timeline"),
new Note("Task", "Review pull request"),
new Note("Reminder", "Update documentation")
];
const result = promptWithNotes('User\'s Notes')`${notes}`;
console.log(result);
Output:
==== User's Notes ====
• Meeting
Discuss project timeline
• Task
Review pull request
• Reminder
Update documentation
==== End of User's Notes ====
Custom Array Formatting:
You override the default array formatter (two line seperation: '\n\n') by providing a custom formatter in the makPrompt for the Array type:
const commaPrompt = makePrompt({
Note: (note) => note.title,
Array: (items, formatter) => items.map(formatter).join(', ') // comma separated items
});
const result = commaPrompt('Note Titles')`
Meeting topics: ${notes}
`;
console.log(result);
Output:
==== Note Titles ====
Meeting topics: Meeting, Task, Reminder
==== End of Note Titles ====
class Task {
constructor(public name: string, public completed: boolean) {}
}
const customPrompt = makePrompt({
Note: (note) => `📝 ${note.title}: ${note.content}`,
Task: (task) => `${task.completed ? '✅' : '⏳'} ${task.name}`
});
const note = new Note("Meeting", "Discuss project timeline");
const task = new Task("Review code", false);
const result = customPrompt('Project Status')`
Current items:
${note}
${task}
`;
console.log(result);
Output:
==== Project Status ====
Current items:
📝 Meeting: Discuss project timeline
⏳ Review code
==== End of Project Status ====
Prompts can be nested and composed together:
const systemPrompt = prompt('System')`
You are a helpful AI assistant.
`;
const userContext = prompt('User Context')`
User: Alice
Task: Help prioritize work
`;
const fullPrompt = prompt()`
${systemPrompt}
${userContext}
Please provide helpful guidance.
`;
console.log(fullPrompt);
Output:
==== System ====
You are a helpful AI assistant.
==== End of System ====
==== User Context ====
User: Alice
Task: Help prioritize work
==== End of User Context ====
Please provide helpful guidance.
The library includes comprehensive tests that demonstrate usage patterns:
import { prompt, makePrompt } from 'llm-prompt-tag';
// Basic usage
const result = prompt('Intro')`
Hello world.
`;
// Output: "==== Intro ====\nHello world.\n==== End of Intro ===="
// With custom formatters
class Note {
constructor(public title: string, public content: string) {}
}
const noteFormatter = (note: Note) => `• ${note.title}\n${note.content}`;
const customPrompt = makePrompt({
Note: noteFormatter
});
const note = new Note("Meeting", "Discuss project timeline");
const formatted = customPrompt('Notes')`${note}`;
// Output: "==== Notes ====\n• Meeting\nDiscuss project timeline\n==== End of Notes ===="
Creates a tagged template literal function for building prompts.
Parameters:
- header (optional): String to use as section header
- condition (optional): Boolean to conditionally render the section (default: true)
Returns: A tagged template function
Creates a customizable prompt function with type-specific formatters.
Parameters:
- formatters: Object with constructor names as keys and formatter functions as values. Special key Array can customize array formatting.
Returns: A prompt function with custom formatting capabilities
Example:
const customPrompt = makePrompt({
Note: (note) => `• ${note.title}\n${note.content}`,
Task: (task) => `[${task.completed ? 'x' : ' '}] ${task.name}`,
Array: (items, formatter) => items.map(formatter).join(' | ') // Custom separator for list items
});
# Install dependencies
npm install
# Build the library
npm run build
# Run tests
npm test
# Development mode with watch
npm run dev
The library consists of two main modules:
Core tagged template implementation with whitespace cleanup and conditional rendering.
Extension mechanism for registering custom object formatters.
MIT
Dave Fowler
GitHub | X / Twitter | ThingsILearned
