How to write a note your AI can't improve

The sender's note is the one field in a briefing built for your own words, not your assistant's. Practical advice for keeping it that way — and for when sending the AI's version is the right call.

4 min read

A RelayLink briefing has one field your assistant did not write for you: the sender's note. It drafts the ask, the context brief, the assumptions, the options — all of it reconstructable from the session. The note is different by design. It's the one place a human is supposed to speak directly, and most people either skip it or let the draft stand in without reading it twice. Here's how to use it for what it's actually for.

Write it before you read the draft

Anchoring is real, and it works against you here specifically. Once you've read a fluent, well-organized draft of "your" note, your own phrasing starts to feel clumsy next to it — not because it's worse, but because the draft arrived first and reads smoother on a first pass. So people default to the assistant's version, not because it says more, but because it already said something and going second feels like extra work.

The fix costs nothing but order. Write your one or two sentences before you ask your assistant to draft the note, or at minimum before you read what it proposed. Compare after, not during — comparing during is how a perfectly good sentence quietly loses to a smoother one that says less.

Keep it to what only you can say

Everything reconstructable from the session belongs in the context brief — your assistant is good at that, and there's no reason to duplicate it in your own words. The note is for the rest: your actual reaction, a piece of history the session doesn't contain, the register the situation actually deserves.

A note that fails this test reads like a smaller version of the context brief: "I've reviewed the numbers and recommend option B, given the timeline." True, and also just the ask restated — your assistant already said as much, in more detail, two fields up. It adds nothing because it contains nothing only you could know.

Compare that with: "This matters to me more than the memo lets on." "I should have raised this last week — that's on me." "I know how this reads; I mean it anyway." None of those are reconstructable from a transcript of you working through the problem with a model. They're not conclusions, they're stances, and a stance is the one thing an assistant can infer but never actually hold on your behalf.

Say the thing you'd be uncomfortable not saying

A useful test for what goes in those two sentences: imagine the recipient reads only the note and nothing else in the briefing. What's the one line you'd want to have come from you personally, rather than from the summary of your reasoning?

It's often the line that feels slightly exposed to write — an apology, an admission that you're not fully sure, an acknowledgment that you're asking for more than the context brief lets on. That discomfort is a decent signal you've found the right sentence. A note that reads exactly like the context brief, just shorter, hasn't done its job.

Sending the AI's wording is a fine outcome

RelayLink's provenance labels exist precisely because both outcomes are legitimate. Write your own words, and the note ships verbatim, human-authored. Read the draft, decide it already says what you meant, send it unchanged, and it ships AI-drafted, approved unchanged. Both are honest. Neither is a failure.

The label is earned, not asserted: it appears only when your final wording differs from anything the assistant drafted, because echoing a draft back is approval, not authorship. That means the label describes what happened — it shouldn't decide what you write. Don't tweak a word or two just to trigger the verbatim tag; a rewrite performed for the metadata isn't any more yours than the draft was. Write first because it gets you your actual sentence, not because a label is watching.

Before you draft anything

Give your assistant the send and it will handle the ask, the context, and the options competently — that part scales well. The note doesn't, and isn't supposed to. It's the one field built to scale with you instead of with the model, which is the whole reason to leave it alone. Write your two sentences first, keep them to what only you could say, and let the label record whatever actually happened.

The note is the last step; most of what makes a draft need rewriting happens earlier, in the session itself — how to brief your assistant so the draft lands covers that half.

To see it on a real briefing, connect your assistant.

Frequently asked questions

What should I write in the sender's note on a briefing?
One or two sentences that only you could produce — your actual reaction, an acknowledgment, a stake in the outcome, or context the session itself doesn't contain. Anything reconstructable from the conversation belongs in the context brief, which your assistant already writes.
Should I write my note before or after reading the AI's draft of it?
Before. Reading a fluent draft first anchors your own wording against it, and your version starts to feel clumsy by comparison even when it says something the draft can't. Write your sentences, then look at what the assistant proposed.
What happens if I just send the AI's version of my note?
It ships labeled "AI-drafted, approved unchanged" rather than "verbatim, human-authored" — echoing a draft back is approval, not authorship. That's an honest, ordinary outcome, not a failure, and it isn't worth a forced rewrite just to change the label.