Feeding Your AI the Docs It's Missing

Claude kept fucking up my Laravel queue jobs. Every time I asked it to set up a queue worker or configure broadcasting, it would give me Laravel 9 syntax. The commands were wrong, the config files were in the wrong place, and I was spending more time fixing its output than I would have spent writing it myself.

The model's training data has a cutoff. It knows Laravel, but it knows the Laravel from whenever its training ended. Laravel moves fast. The queue system changed. Broadcasting changed. The whole application structure changed in Laravel 11.

I needed to feed it the current docs.

The First Attempt: Repomix

Found a tool called repomix that compresses a whole repo into a single text file. Perfect for uploading to Claude's Projects feature.

Grabbed the Laravel docs repo, ran repomix, tried to upload it.

Problem: it used 450% of the project's knowledge capacity. The Laravel docs are massive. Too massive.

The Solution: Delta Condensing

I only needed Claude to know what had changed since its training cutoff. It already knows Laravel 10. I just needed to teach it Laravel 11+ differences.

So I wrote a command that:

1. Takes the full Laravel docs
2. Sends each section to GPT-4 (cheaper than Claude for this)
3. Asks "what's new or changed from Laravel 10?"
4. Keeps only the delta

private const PROMPT = <<<'EOT'
You are helping create training data about Laravel 11+. You already know Laravel up to version 10.

Keep ONLY information that is:
- New in Laravel 11+
- Changed from Laravel 10
- Non-obvious implementation details specific to Laravel 11+

Remove everything that:
- Was already present in Laravel 10
- Is basic PHP or web development knowledge
- Is introductory or explanatory text

Format as concise bullet points.
If there's nothing new, respond with "NO_NEW_CONTENT".
EOT;

The full command runs through each markdown file in the docs, condenses it, and outputs a single file of just the changes.

The Result

The full Laravel docs: 450%+ of project capacity (unusable) The condensed changes: 12% of project capacity (perfect)

Now Claude knows about the new application structure in Laravel 11, the streamlined bootstrap/app.php setup, the updated queue configuration, health check routes, and all the other shit that changed.

How I Use It Now

I don't actually use Claude Projects much anymore. With Claude Code, I use a CLAUDE.md file in my project root that contains:

1. Project-specific patterns and conventions
2. My preferred stack versions
3. Links to or excerpts from current documentation
4. Things Claude consistently gets wrong that I want it to remember

The CLAUDE.md approach is better for active development because it lives in the repo and travels with the project. But the doc-condensing technique still works great for building that initial knowledge base.

The Command

If you want to do this yourself, the basic flow is:

// Pseudocode
foreach ($docsFiles as $file) {
    $condensed = $llm->ask(PROMPT . $file->contents());

    if ($condensed !== 'NO_NEW_CONTENT') {
        $output->append($condensed);
    }
}

Full gist here: gist.github.com/StuMason/...

You can adapt this for any framework or library that's moved faster than your AI's training data. Works great for Inertia, Livewire, Tailwind - anything with active development and good docs.