What changes when the agent knows you

The problem with generic newsletters

Most newsletter content is written once and shipped to thousands. The author picks the subjects that matter to "the average reader" — and for any individual subscriber, half the content is irrelevant.

In a research project on agent-written email, you'd think the natural answer is: just write more newsletters. One per niche. One per persona. One per segment. But that's the same broadcast model with smaller buckets.

The actual move is one newsletter — written differently for every single reader.

What we built

When a reader signs up, they enter their own domain. The context agent immediately reads their homepage, pricing, features, about, and blog. It drafts five markdown docs:

• positioning.md — who they are, what they sell, who they serve
• icp.md — ideal customer profile, inferred from the site
• pricing.md — tiers, value metric, free vs paid
• voice.md — tone rules, banned words, a quote that exemplifies their style
• differentiators.md — what they emphasise that rivals don't

These docs are editable. The reader can rewrite any of them. Locked docs survive the weekly auto-refresh; unlocked ones get re-drafted as the site changes.

Then every Monday, when the editor agent composes the weekly brief, those five docs go into the prompt before the competitor signals. Every takeaway is anchored to who the reader actually is.

What changed

Before context docs, the editor's "what you should do" lines were the predictable thing AI writes when it has no anchor: "Review their pricing page. Audit your messaging. Stay informed about market trends." Useless.

After context docs, the same model started writing things like:

Stripe just shifted to outcome-pricing at $0.99 per result. Your pricing hero still leads with "starts at $99/seat" — your seat model now reads as legacy. Replace the lead line with outcome framing by Friday.

That's not a category of advice. That's a specific instruction grounded in what the rival did and what the reader currently says about themselves. The model didn't get smarter. It got context.

The cost is laughable

Each context-doc bootstrap costs about $0.01-0.02 on Claude Haiku. Weekly refresh is the same. At one thousand readers, ~$15-30/month for the entire personalisation layer.

Compared to a human marketer writing personalised content for one company — at agency rates, that's two hours of work, every week, for one subscriber. Times a thousand.

This is the headline finding so far: the marginal cost of personalisation collapses to near-zero when the agent has structured context about each reader.

What's next

We're testing three things:

1. Can the agent identify themes across multiple subscribers' competitor sets — patterns of moves only visible at the population level, surfaced back into individual briefs
2. Can the voice doc enforce voice consistency across drafts — automatically checking the editor agent's output against banned phrases
3. Can the context agent infer changes to the reader's own positioning week-over-week — diff against their site and surface "your hero changed; here's how the digest will adapt"

If you're building agent-composed content, the lesson so far is unsubtle. Don't ask the agent what to write. Ask the reader who they are. Then ask the agent.