More context won't save your prompt

Hi friends!

Somewhere in your notes is a prompt that works.

You built it over a few months of trial and error. It has a role line, a clear task, probably a format instruction, maybe a “keep it under 400 words.”

You paste it in, swap the topic, and out comes something usable. It used to be great, but now it’s reliable.

And you cannot figure out what separates your B from the A you got at the beginning, or by accident in one of your chats .

So you do the reasonable thing.

You add more. More instructions, more examples, a longer role description, three extra bullet points about your brand, one more paragraph of context.

And the output gets... not better? And sometimes actively worse. And you want to throw your laptop out the window. Oh I know how THAT feels.

The reason is simple once you see it.

Claude does NOT read everything you paste and quietly pick out the good parts.

It weights all of it.

Everything you hand it, it treats as a thing you want more of.

Let me show you with the example every food blogger reaches for first.

You want Claude to sound like you, so you paste in your last ten newsletters.

Feels obviously right. More me, more voice, better match.

Except three of those ten were sends you did not love.

One was a heavy roundup that does not sound like you at all.

Two are from before you found your voice.

You just handed Claude ten things and told it, without meaning to, that all ten are equally “you.”

So it averaged them. It gave you back the mean of your archive, and the mean of anyone’s archive is mush.

The move is not more. The move is CURATION.

Pull the two sends that sound the most like where you are going, not where you have been.

Then, and this is the step people skip, tell Claude WHY those two.

“Use these. The opener in the first one is exactly my voice, a little dry, gets to the point fast. The second one rambles in the middle and I want you to know I like that it rambles.”

You have not given it more. You have given it a filter.

You have done the one thing it cannot do for you, which is decide which version of you is the real one.

I would tattoo this on the inside of my wrist, metaphorically.

It is in no prompt template, because it is not a prompt at all.

Ask for three drafts, never one.

Read them and tell Claude the truth, out loud, the way you would talk to a writer you were mentoring.

This opener is trying too hard.

This one sounds like a brand and not a person.

This middle paragraph is the only honest thing on the page, do more of that.

Then send it back to write again, carrying what you just said.

Watch what happens on the second pass.

It is not just better. It is better in the EXACT direction you pointed.

You did not write a new prompt. You edited, it adjusted to you, and that loop is the whole difference between a tool and a collaborator.

Here is why this beats any template.

Your taste is THE asset. Not your prompt.

A food blogger who knows, in her gut, that this sentence is honest and that one is performing has the one thing the model will never have and that every competitor on the same subscription is failing to use.

The template crowd is busy perfecting the input. You should be running the loop.

So this week, do not add a single thing to your prompt.

Take the B-grade reliable one you already have and cut its context down to the two best examples with a sentence each on why they made the cut.

Then run it, ask for three, and edit like you mean it.

Say what is wrong. Send it back. It is just like working with a brand new VA. They might be smart, but they don’t know how YOU work just yet. So teach them.

You WILL get your A. And you will notice you got there not by describing yourself more completely, but by choosing what to show and reacting to what came back. That is not a form you fill in. That is a relationship you run.

And if you’re thinking, “Okay Shruthi, I like this, but I am not hunting for the two right examples and running a loop from scratch every time I open a chat”

And that’s fair. You shouldn't have to.

There is a way to do the choosing once and mostly forget it. That is where we go next. Bonus newsletter dropping tomorrow :)