Public relations

What Sam Polstein keeps in a folder called good writing

Sam Polstein puts AI output quality at 90% decided before the prompt. What is in his context library, why it is markdown, and why he owns the files.

The highest-leverage thing a communications team can build right now is not a prompt library. It is a context library: client material, house process and examples of good writing, kept as plain markdown files any model can read. Sam Polstein, who leads communications and content for deep tech startups at HAUS, puts the number at 90 percent.

"I think by the time the prompt has started, the output is already 90% of the way there. It comes down, in my experience, to context. I think about context engineering a lot more than I think about prompt engineering."

What is context engineering in communications?

Context engineering is the practice of curating the material a model works from — client calls, decks, technical papers, house process, examples of writing you consider good — instead of optimising the prompt. In Sam Polstein's account, quality is decided before anything is typed into a chat window.

He argues his field needs it more than most: deep technical subject matter, a history of approaches that worked and did not, and a job that consists of pulling something complex out for a more general audience. Curation includes telling the model what not to reach for: "Giving it the right context and making sure that it's only searching the web and using the web and using its general knowledge sparingly and in the right context."

What actually goes in the library?

Four kinds of material, in Polstein's setup, all stored as markdown files in Obsidian and pointed at with Claude Cowork:

Client context. "Claude Cowork can jump in and when I direct it at a certain folder, can see all the context for a given client, for a given process."
House process. The steps a deliverable is meant to pass through.
Examples of good writing. A folder literally called good writing.
Tropes to avoid. A list the model checks its output against before it hands anything back.

The last two are the layer most teams skip, because they are not about any client at all.

"There's an additional layer of context that we're building just internally on our own, which is: how do you become a good writer? I've got a context folder on my computer that is just called good writing. And I fill it with things that both I think are good writing and also a list of tropes to avoid, things to make sure that Claude is checking before it gives me anything."

He treats maintaining it as building, not admin.

"That in and of itself is something that's ever evolving and that I kind of think of as building a tool in and of itself. How can I create a context folder that will make my Claude outputs better?"

Why does it have to be files you own?

Because the model you use this year is probably not the model you use next year, and the context has to survive the change. Polstein is deliberate about where his lives.

"I want to make sure that the context is mine, it's not Claude's ... I keep it in Google Drive just to make sure it's backed up, but at the end of the day it's just a bunch of folders of markdown files, plain text files."

The reason is portability.

"This way, if Claude starts becoming really bad or raises their prices a ton, we can use another tool. Or if ChatGPT releases a new model that's just an incredible writer and we just have to use this new model, we can. And it's just a question of switching over, and the switching costs don't become so great."

He has watched the alternative walk into his agency.

"I've talked to folks who have come into the agency and they've been like, well, I have really trained ChatGPT — it knows my writing, it knows who I am, I've spent a lot of work. And then it's hard to pull that out and bring it into another service."

Work stored inside one product cannot leave with you. Plain text can.

Does it matter how bluntly you talk to it?

Only in the sense that precision matters. Polstein describes his own style unflatteringly and then, on reflection, more accurately.

"I'm kind of condescending, frankly. Like I'm very specific. Definitely a micromanager in terms of: I want you to look in this file and pull out this thing, because this was where we talked about the strategy that I want you to focus on."

Galina Fendikevich pushed back on the word — that reads as direct, not condescending — and Polstein agreed, with an analogy worth keeping.

"If I ask somebody, a report of mine, in person to do something, they're picking up on all my body language ... Claude is basically that, even more so. It has no idea, it's just the words that I'm inputting. There's no metadata to that prompt that's like: how does Sam actually feel about this?"

Directness is not rudeness. It is compensating for a channel with no body language in it.

What to do this week

Start with the folder that is not about a client. Put five pieces of writing you would be happy to publish under your own name into one place, then write a second file listing the phrasings you never want to see again. Point the model at both before it drafts, and name the specific file and passage you want it to work from.

Then move all of it somewhere you own — plain markdown, backed up, no product in the middle. If you had to change models tomorrow, how much of what makes your work good would leave with the old one?