Why mega prompts don't fix AI output

Yo, another email coming at you from yet another city.

Last summer I was building a tonne of Custom GPTs.

And, honestly, I got pretty good at uilding them, and getting results from them...

... I made multiple of them available to the folks in GM+.
... Ran some as their own LTOs to generate new sales and leads
... Saved hours of creation time by getting the V1 draft out the way

When agents came in though, I tried to just port everything over.

I would write these MASSIVE prompt and istrcution files. Like 2800 words for one, including every rule and specifci I could think of.

My thought was, if I cover all the ground, there's no way it could mess it up.

Most folks who care about getting better AI output have tried something like thisalready...

... Downloaded a prompt pack from Reddit.
... Pasted a CLAUDE.md template from someone's GitHub.
... Saved a tone-of-voice doc cribbed from a Cursor user on Twitter.
... Added a couple of skills they'd seen demoed somewhere.

And often, on the first run, these things work well.

But, over time, the output degraded and you ended up with something that's back to being em dash-filled generic BS.

The ToV file stops firing. Claude ignores half the rules. You add more rules. They contradict each other. You strip some out. Things get worse.

I've done all of this, several times. I have a graveyard of instruction drafts in old project folders.

And it's shown me how AI actually works and what you need to do to make sure it's actuaslly helping you, not hindering you.

Essentially, Cluade or whatever LLM skims the files as a human would a blog post.

Most people don't read word-for-word, neither does Claude.

I think it's mainly to do with the context window.

If you overload it, it gets confused. So it cherry picks what it should be looking at.

And that leads to really inconsistent results.

What I've found to be a universal truth with AI is that, what's in your files is important, but the structure and relations between those files is more so.

if you just dump everything in one file, it's never gonna work.

But having the right hierarchy, file relations, and handoffs is what helps you get more done, at less cost, and with higher accuracy.

When all you're doing is adding a new line like "don't do em dashes" to the general prompt, it's more noise to the system.

Most of us are just adding components to our instructions and file bases, when you should be designing the wiring.

That's the part I want to show you. What it looks like when the architecture is right. Coming in the next one.

Pete "graveyard of CLAUDE.md drafts" Boyle