README.md

Blog post code generator tool

Introduction

This is an example of using Metaphor to build a tool that is used to create blog post entries for the m6r.ai website. The aim is to take a markdown file, blog.md and have a large language model (LLM) generate the same content as TypeScript code.

The website runs as a single page application, that uses a custom virtual DOM implementation. You don't need to worry about the details, but essentially this means each blog post is encoded as an HTML-like tree structure using a series of nested functions.

The example showcases a few important elements:

• It shows how to use an Action: section to describe what you want the LLM to do.
• It has a Context: section that embeds the blog.md file that is to be used as the basis for the output
• The Context: section also has a blank template for what the output file should be.
• The Context: section includes 2 examples of other blog posts to demonstrate to the LLM what blog posts should look like.

Using m6rc

To generate a large context prompt (LCP), use the following:

python3 ../m6rc/src/m6rc/m6rc.py blog-editor.m6r -o out.lcp

In this case the path to m6rc.py is ../m6rc/src/m6rc/ but you can replace this with whatever your path needs to be.

Once run it will generate a file, out.lcp. If you copy this to the interactive window of the LLM it can generate a response.

Outputs

Our example asked for this text to be transformed:

This is a test to see if I can make this work

## We've got headings

And we've got text.

This is a really long line.  It will be *interesting* to see how this gets handled and whether the LLM can split things
across several lines of source code?

### Here's a different heading

More text

- A bullet
- A second one
- And a third one

Isn't this fun?

Both ChatGPT 4o and Claude 3.5 Sonnet give us the correct form of output. The only difference is ChatGPT uses an em tag, where Claude uses an i tag. According to the original Markdown spec (see: daringfireball.net/projects/markdown) this should probably be em, but a minor addition to the Metaphor description can make this the same for all implementations (you might try it as an exmple).

ChatGPT 4o

ChatGPT generates the following:

import {h, VElement} from '../../lib/dvdi';
import {BlogPost} from '../BlogPost';

function blogArticle_000004(): VElement[] {
    return [
        h('p', {},
            'This is a test to see if I can make this work'
        ),
        h('section', {},
            h('h2', {}, 'We\'ve got headings'),
            h('p', {},
                'And we\'ve got text.'
            ),
            h('p', {},
                'This is a really long line.  It will be ',
                h('em', {}, 'interesting'),
                ' to see how this gets handled and whether the LLM can split things across several lines of ' +
                'source code?'
            )
        ),
        h('section', {},
            h('h3', {}, 'Here\'s a different heading'),
            h('p', {},
                'More text'
            ),
            h('ul', {},
                h('li', {}, 'A bullet'),
                h('li', {}, 'A second one'),
                h('li', {}, 'And a third one')
            ),
            h('p', {},
                'Isn\'t this fun?'
            )
        )
    ];
}

export const blogPost_000004 = new BlogPost(
    'Title',
    'Dave Hudson',
    '2024-11-03',
    '/blog/000004-Blog-post',
    'Summary.',
    null,
    blogArticle_000004
);

Claude 3.5 Sonnet

Claude gives almost exactly the same response, except for using an i element instead of an em:

import {h, VElement} from '../../lib/dvdi';
import {BlogPost} from '../BlogPost';

function blogArticle_000004(): VElement[] {
    return [
        h('p', {},
            'This is a test to see if I can make this work'
        ),
        h('section', {},
            h('h2', {}, 'We\'ve got headings'),
            h('p', {},
                'And we\'ve got text.'
            ),
            h('p', {},
                'This is a really long line. It will be ',
                h('i', {}, 'interesting'),
                ' to see how this gets handled and whether the LLM can split things across several lines of source code?'
            )
        ),
        h('section', {},
            h('h3', {}, 'Here\'s a different heading'),
            h('p', {},
                'More text'
            ),
            h('ul', {},
                h('li', {}, 'A bullet'),
                h('li', {}, 'A second one'),
                h('li', {}, 'And a third one')
            ),
            h('p', {},
                'Isn\'t this fun?'
            )
        )
    ];
}

export const blogPost_000004 = new BlogPost(
    'Title',
    'Dave Hudson',
    '2024-11-03',
    '/blog/000004-Blog-post',
    'Summary.',
    null,
    blogArticle_000004
);

Iteration

One of the advantages of using Metaphor is the ability to evolve requirements and pull in external context. If you check the git history you'll see that the blog structure was originally slightly more complex, but was later simplified. The ability to capture this change in requirements and the changed Action: made updating the LCP for the LLM much simpler.