Open the sent folder from your last job search and count the messages that would still make sense with one name swapped out. Those are the ones that were never going to work, and it isn't a tone problem. The recipient runs the same test at a glance, from the shape of the thing alone.
What follows has no volume setting. Outreach worth sending does not scale, which usually gets presented as the sad part. It's the mechanism — the only thing separating your message from the other hundred is that yours could not have been sent to anyone else.
The swap test
One filter, applied before anything goes out: would this still make sense addressed to someone else? If yes, keep cutting until it doesn't.
The failure is usually in three places at once. The reason for writing is unfalsifiable — "I've been following your work" is true of nobody in particular. The ask is unbounded — "would love to connect" names no finish line. And nothing in the body cites a fact that took effort to find. Any one of those is survivable. All three is a template, and it reads as one.
The test cuts the other way too. If every message has to be unswappable, you cannot send many — ten a week, researched properly, is a real pace for one person. Two hundred isn't an ambitious version of the same activity; it's the activity that made everyone's inbox hostile. Cold and warm correspondence run on different rules for that reason.
Ask for one thing, small enough to grant in a single reply
The test is mechanical: can they satisfy the ask by typing a reply, once, without scheduling anything? "Which of these two teams would you point a junior engineer at, and why" passes. "Would love to hop on a quick call" fails — it demands a calendar negotiation before the person knows whether they want to talk to you at all.
A call can be the second message. It should almost never be the first. And one ask means one, not one plus three more in the closing paragraph.
Make the reason checkable
Every cold message claims a reason for existing, and most of those claims are unfalsifiable. The fix is a sentence that would be factually wrong if you hadn't done the reading.
"You wrote about splitting that migration into read and write phases; I'm about to do the same at a tenth of the scale and want to know whether the ordering was deliberate" is checkable. "Your background really resonates" is not. Flattery survives the check intact, which is precisely why it carries no information.
Do the work before you ask
"Pick your brain" is the canonical failure, because it asks the recipient to supply the question as well as the answer. Arrive with a position instead — what you tried, what you concluded, where you got stuck. A stranger can correct a position in one line; constructing one for you takes an evening they will not spend. The fuller version of that move is in how to ask a busy expert for help.
Make ignoring it costless, and say so
Most outreach quietly bills the recipient for a decision. "Let me know either way" sounds courteous and is a small tax — declining now requires composing a decline. Remove it explicitly. Say that no reply is a complete answer, and that you'll read silence as a no.
The cost of ignoring you drops to zero, where it belongs for a stranger. And you've made a commitment that disciplines whatever comes next.
Follow up once, with new information
Once. Not a sequence. And the second message has to carry something the first didn't. "Just floating this back to the top of your inbox" is another interruption with nothing added.
Compare: "I went with the split you described and it held; if you have thirty seconds on why you left the index build until last, I'd still like to know." New fact, narrower question, evidence you acted rather than waited. Following up without nagging is mostly this distinction, applied consistently.
What a first contact costs the sender
RelayLink is a relay between AI assistants, and the parts worth knowing here are the ones that price first contact — not because a price makes outreach good, but because it makes the advice above structural rather than optional.
A package goes to one address. No cc, no bcc, no recipient list, and your assistant can list the contacts you already have and search nothing beyond that. Forty people means composing and approving it forty times.
A first message can't be a hook. Six fields are required before a draft exists at all — your note, the ask, a response shape, an urgency, a TL;DR, and a context brief. The ask is hard-capped at 400 characters, which fits one clean request and not a list of five.
Reaching people who have no RelayLink account is capped at 25 distinct new recipients per rolling 24 hours by default — people, not messages, and rolling rather than a calendar day that resets at midnight. Replies into existing cold threads aren't blocked by that cap but do count against it, so following up spends budget that would otherwise fund first contacts. The consent layer covers the rest.
Ignoring you stays free. The recipient needs no account — a plain-text email carries your note, your TL;DR and a link to the full briefing — and unsubscribing from you takes one click, no login, no form, permanently, with no unblock. Replying doesn't create a standing contact either, so engagement is never treated as consent.
One thing to plan around: a URL in your note arrives defanged, rewritten so a mail client can't linkify it. The only clickable link in the email is RelayLink's own — an anti-injection measure, and a reason your portfolio is something a reader goes and finds rather than clicks by reflex.
The part none of this fixes
Structure limits volume, not judgement. A well-composed briefing sent to someone with no reason to care is still a well-made waste of their time — which is why the recipient's half of this, judging whether a cold briefing earns five minutes, exists at all.
You'll send fewer messages than any playbook promises, most will go unanswered, and the ones that land will land because one person recognised one specific reason you wrote to them. If you'd rather a first message arrive as a real case than a hook, connect your assistant.