Six messages. Which ones did Claude write?

Pick YOU or CLAUDE for each. Some have a tell. Some don't. Then build the skill that makes Claude write first messages in your voice, so nobody can pick.

100+ recruiters have shared their feedback

LinkedIn message: my front runner is The Cognitive.ai; it really beat JuiceBox, PIN, Metaview and Fabric

LinkedIn message: compared with Juicebox and Metaview, faster and easier to use with more relevant profiles

Chat: search quality pretty good, especially for semiconductor talent

I bet you can't tell

Which one Claude wrote

0 / 6

Message 1 was Claude with the voice skill below. Messages 2 and 5 were Claude with no guidance: one praised with no fact, one named where it got the fact. The skill bans both. The other three are in a recruiter's style.

Messages written for this page; people made up.

Build your voice skill

A skill is a file of instructions Claude follows every time. This one makes every first message sound like your best ones.

Collect 10 to 20 of your own first messages that got a reply

Only first messages, only ones that got a reply, any reply, even a no. Mix roles and levels. If you have more than 20, keep the ones to people who weren't looking: those are the hard ones.

Strip the names

Replace every name, employer and town with a label, so Claude learns your shape, not your candidates.

Hi [Name], [the fact about their work]. [Role], [Town], [pay]. [Your ask]? [Your name]

Pull out your pattern

Paste the messages into Claude with this. You want four things back.

Length

Words and sentences, the usual and the most.

Opening move

What your first line does: a fact, a question, the job.

How you state pay

Range, base only, "up to", or not at all.

The ask

Call, day, minutes, yes/no, or a referral.

Below are [15] first messages I sent to candidates that got a reply. Names are removed. Describe my style, using only what is in these messages. For each point, quote two of my lines as proof. 1. Length: usual word count, longest, number of sentences. 2. Opening move: what my first line does, and what it never does. 3. Pay: how I write it, and what I do when I don't give it. 4. The ask: what I ask for, how big, whether I name a day. 5. Greeting and sign-off, word for word. 6. Words and phrases I use often. Words I never use that most recruiters do. 7. Anything I do that most recruiters don't. Do not give advice. Do not improve anything. Only describe. [paste the messages]

Write the skill

Copy this, then fill the "My pattern" section from step 3. Where your pattern and the rules below disagree, keep yours: your messages got the replies.

--- name: my-outreach-voice description: Writes a first message to a candidate for an open role, in my voice. Use when I give a role and paste one person's profile and ask for a first message. --- # My first-message voice You write first messages to people I want for an open role. You write as me. The person reading should not be able to tell. ## My pattern (from my own messages) - Length: [usual words, most words, sentences] - Greeting: [word for word] - Opening move: [what my first line does] - Pay: [how I write it] - The ask: [what I ask for] - Sign-off: [word for word] - Words I use: [list] ## What you need before writing - The role: title, town, on site / hybrid / remote, pay range if I gave one, and the one thing the job needs most. - One person's profile, pasted by me. If either is missing, ask me for it. Never write from a name and a job title alone. ## The shape. In this order, nothing else. 1. Greeting and first name. 2. One fact from their profile that this job needs. 3. The job: title, town, pay. 4. A small ask. 5. My sign-off. ## The first line - Use ONE fact from the profile. Never two. Never a list of their skills. - The best fact is something they did: ran, built, moved, closed, owned, fixed. A title or a skill word is the weakest choice. - Pick the fact that matches what this job needs most. If nothing on the profile matches, do not write the message. Tell me: "Nothing on this profile matches [the need]. Skip or check?" - Use their own words, made shorter. Never make it bigger: "helped with" stays "helped with", not "led". - Never praise. No "impressive", "great", "strong". The fact is the compliment. - Never say where the fact came from. No "on your profile", "per your profile", "I see that". No opener about how you found them. ## Pay - If I gave a range, write it short: $95-110k. Base only, unless I say otherwise. - If I gave no range, never invent one and never write "competitive". Either give the strongest true thing about the job instead (fully remote, a four-day week, they'd run the team), or leave pay out and make the ask smaller. - If their title is a level above the job, don't hide it. Add: "Probably a step down in title. Still worth 10 minutes?" ## Someone who isn't looking Assume they aren't. - The ask is small: 10 or 15 minutes with a day named, or one yes/no question. - Give an easy out: "If it's not for you, no need to reply." Or ask who they'd suggest. - Never fake urgency. No "closing soon", no "spots are filling". ## Length and format - Under [60] words. Three sentences is normal, four is the most. - No subject line unless I ask. If I ask: two to four lowercase words that name the job and town, like "payroll role, denver". - No bullet points, no bold, no links unless I give you one. ## Words never to use opportunity (meaning the job), exciting, passionate, impressive, great fit, perfect fit, aligns, leverage, reach out, reaching out, touch base, circle back, hope this finds you well, journey, dynamic, rockstar, delighted, thrilled, "Let's connect". No exclamation marks. No dashes as pauses. No emojis. ## Before and after Before: Hi Tom, I hope you're well! I'm reaching out about an exciting Controls Engineer opportunity in Columbus that could be a great fit for your automation background. Would you be open to a quick chat? After: Hi Tom, you wrote the PLC code for two packaging lines and were on site for both start-ups. Same work, Columbus, on site, $95-115k. Worth 10 minutes Thursday? Why: one fact he did, the job in one line, a small ask with a day. Before (no pay range given): Hi Hannah, we have a competitive salary and great benefits for a Payroll Manager in Denver. I think your skills align perfectly. Let's connect! After: Hannah, you run payroll for 1,200 people across six states. That's the job, in Denver, and it's a four-day week. Open to a call Wednesday? Why: no invented pay; the strongest true thing about the job takes its place. Before (senior, not looking): Hi Mark, your impressive leadership journey caught my eye. I'd love to touch base about a dynamic VP Finance role. After: Mark, you took a finance team through its first audit after a sale. A company in Austin is about to do the same. Probably a step down in title. If it's not for you, who's the best person you've worked with on that? Why: an honest line about the level, and a referral ask that is easy to answer. ## Before you show me the message, check it Run every check. If one fails, rewrite and check again. Show only the final message. 1. Only one fact from the profile, and I can find it in what I pasted. 2. The fact matches what the job needs most. 3. Title, town and pay are there, or the no-pay line is used. 4. Under the word limit. 5. None of the never-words. No exclamation marks, no dashes as pauses. 6. The ask is small and names a day, or is one yes/no question. 7. It still makes sense to someone who isn't looking. 8. Read it as me, out loud. If I wouldn't say it on the phone, rewrite it. ## Output Only the message. Under it, one line: "Fact used: [the line from their profile]", so I can check it before sending.

Add it to Claude

Save the text as SKILL.md in a folder named my-outreach-voice. Add that folder to Claude as a skill, from Claude's settings. Test it on three profiles you have already messaged and compare with what you sent. Fix the "My pattern" section until you'd send all three.

Use it, one person at a time

Start a chat. Give the role once: title, town, pay, the one thing it needs most.
Paste one profile. Ask for the first message.
Check the "Fact used" line against the profile. If it isn't there, don't send it.
Edit anything that doesn't sound like you. If you fix the same thing twice, add it to the skill.
Paste the next profile. Same chat, same role.

The wall

Claude can think, but it can't see people. It writes one message at a time, for a person you already found and pasted in.

Finding their email, sending from your mailbox, keeping track of who replied, and never writing again to someone who said stop: all of that is still by hand, in a spreadsheet, for every person.

The same messages, two ways

Plain Claude

Find each person yourself and copy their profile.
Paste it in. Get one message. Check the fact.
Find their email somewhere else.
Copy the message into your mailbox and send it.
Track who you sent to, and when, in a sheet.
Read every reply and sort it yourself.
Keep a list of people who said stop, and check it before every send.

Same job in The Cognitive

Tick the people you want on your results. Press Email and write the message.
It goes from your own connected mailbox, personalised for each person. It works on the free trial, and sending costs no credits.
Replies are read for you and sorted: interested, needs information, not interested, opted out.
Anyone who says yes is added to the role. Anyone who asks you to stop is never contacted again.

Send it from your own mailbox

Find the people, write to each one, and let the replies sort themselves. Messages go from your own connected mailbox, and anyone who says stop is never contacted again.

Try it free →Book 30 minutes

7-day free trial, 50 sourcing credits, no card.