Overnight sourcing · the setup
Claude sources while you sleep.
One scheduled task in Claude, with your searches written in the prompt. Every night it picks each search up where it left off, brings back only people you haven't seen, ranks them and drafts the first messages. You wake up to a 15-minute review.
Set it up in 10 minutes
You need a paid Claude plan (Pro, Max, Team or Enterprise) for scheduled tasks, and a free The Cognitive account for the search.
Create your free The Cognitive account
This is the search Claude runs every night, across real profiles, with the line from each profile that made the person a match. Nothing to set up inside it: your searches live in the prompt, and it remembers every search you've run.
7 days free, 100 search credits, no card.
Add The Cognitive to Claude
- In Claude, open Customize, then Connectors.
- Click +, then Add custom connector.
- Paste
https://app.thecognitive.io/mcp - Click Connect, sign in, and approve.
On a Team plan, an Owner adds it once under Organization settings, Connectors. Then each person connects.
Schedule the task
- Open the Claude desktop app and go to Cowork.
- Click Scheduled in the left sidebar, then New task.
- Choose Set up manually, paste the nightly prompt below, and pick Daily at 2:00 AM.
It runs even when your computer is asleep or the app is closed. If you use Claude Code, the same prompt works as a routine at claude.ai/code/routines (a research preview).
Test it once before you sleep
Paste the prompt into a normal Claude chat and let it run. Then run it a second time. Check three things: the second run brought back only new people, it skipped anyone you passed on, and it stopped before revealing any contact or sending anything.
If a search returns junk, fix that search's "Leave out" words in the prompt. See the table below.
The nightly prompt
Paste it as the task's instructions. Replace the 3 searches with your open roles, and change the cap and the Slack channel.
You are my overnight sourcer. Use The Cognitive. My searches (edit these, one per open role): 1. GTM Engineer, 3+ years, remote US. Must show building outbound or data workflows (APIs, enrichment, automation). Comes from B2B SaaS. Leave out: pre-sales and solutions engineers, marketing ops admins, GTM strategy roles. 2. Mid-market Account Executive, 3+ years closing new business, New York, hybrid. Comes from B2B SaaS. Leave out: SDRs, renewal-only account managers, channel roles. 3. RevOps Manager, 4+ years, remote US. Must show forecasting, territories or comp plans. Leave out: revenue accounting, billing, CRM admin only. For each search: a. If I've run this search before, continue that same search and go deeper, so only people I haven't been shown come back. If not, start it. b. Stop at 10 new people per search tonight. c. Skip anyone I've already shortlisted or passed on. d. Judge each new person against the search: Strong = every must-have is shown on their profile and they're in the right place. Possible = must-haves shown or strongly implied, or the level is one step off. Drop everyone else. For each Strong or Possible person, quote the one line from their profile that decided it. e. Shortlist the Strong people. f. Draft a first message for each Strong person, under 90 words: line 1, one specific thing they did, from their profile; line 2, the role, the place and the pay range; line 3, one easy question. Never open with "I saw" or "I noticed". Do not reveal any email or phone number. Do not send any message. Do not invite or reject anyone. Do not buy credits. Those are mine to do in the morning. If a search returns more than half Drops, stop it and tell me the most common reason, so I can fix that search's wording. Morning summary: - Per search: new people, Strong, Possible, credits used - The top 3 people overall, with their quoted line - The drafts, ready to copy - Any search I should reword, and why If Slack is connected, post it to #sourcing. Otherwise leave it here.
"Continue that same search." The Cognitive remembers each search. Going deeper on it brings back only people below the ones you've seen, so you never pay twice for the same face. A brand-new search every night would start from the top again.
A nightly cap. Every person shown costs 1 search credit. 10 per search keeps a 3-search night at 30 credits or less.
The red line. A scheduled task can use every action of a connected app without asking you. In review mode the prompt forbids contacts and sending. Want those too? Swap in the autopilot block below.
The "more than half Drops" rule. A badly worded search burns credits every night. The task stops it and tells you why instead.
Drafts, not sends. You approve every first message. Claude does the 2am part, you do the judgment.
Autopilot: contacts and outreach too
Want the first message out before you wake up? Replace the red "Do not" line in the prompt with this block. Run review mode for a week first, so you trust what it picks.
AUTOPILOT (replaces the "Do not reveal..." line) For each search, take tonight's Strong people, at most 5: 1. Find each person's work email. If none is found, skip them. For searches marked (SMS), find their mobile number instead. 2. If this search has no role in The Cognitive yet, set one up: title from the search, the search as its description, hiring manager: you@yourcompany.com 3. Start an outreach sequence for these people on that role: email, or text for searches marked (SMS), with the draft you wrote as the first message. Never contact anyone I passed on, anyone already messaged for this role, or anyone at a company in that search's Leave out. Never buy credits. If credits run low, stop and tell me. Never invite, move or reject anyone. Add to the morning summary: who was emailed or texted, for which search, and the credits used on contacts.
Contacts. A work email is 4 credits, a mobile number 12. Nothing is charged when nothing is found, and nobody is charged twice.
Sequences. Email or text, with follow-ups on a schedule that stop the moment someone replies, on either channel. Positive replies are read and the person is added to the role, so you start the day with yeses. Sending costs no credits.
Plans. Email lookups work in the free trial. Phone numbers and sending sequences need the Pro plan (or Team for sequences).
Real people, no one watching. A scheduled task acts without asking you, so these lines send real messages at night. Keep the cap at 5, keep "Strong only", and read the first week's sends every morning.
Texting. Texting people who haven't agreed to it is restricted in many places, in the US under the TCPA. Check your rules before you mark a search (SMS).
What to leave out, so you don't wake up to junk
Write these into each search in the prompt, after "Leave out:". The task reads them every night.
| Role | The look-alikes you'll get | Add to that search |
|---|---|---|
| GTM Engineer | Sales engineers, solutions consultants, marketing ops admins, and people with "GTM" in a title that means go-to-market strategy, not building. | Leave out: pre-sales and solutions engineers, marketing ops admins, GTM strategy and enablement roles. Must show building outbound or data workflows (APIs, enrichment, automation). |
| Account Executive | SDRs titled "Account Executive" at small firms, account managers who only renew, and channel or partner managers. | Leave out: SDR/BDR, account managers on renewals only, partner and channel roles. Must show a quota on new business. |
| RevOps | Sales ops coordinators, CRM admins, and finance revenue accountants (same word, different job). | Leave out: revenue accounting, billing, coordinators, CRM admin only. Must show forecasting, territory or comp plan work. |
| Every role | Your clients' staff, people you placed, and people at your own company. | Leave out companies: [your clients], [your company]. Leave out anyone you placed in the last 12 months. |
The 15-minute morning review
Same order every day, so it stays at 15 minutes.
Read the flags first
Any search the task stopped? Reword it in the prompt now, so tonight's run is clean.
Check the Strong list
Open each profile. Pass on anyone who isn't right, so Claude never shows them again. That's how the list gets sharper each night.
Edit and send the drafts
Review mode: find contacts only for the people you'll message today, fix one word in each draft, send. Autopilot: read what went out and answer the replies.
Note one number
Strong per search. Two nights in a row under 2 means widen. More than half drops means tighten. See below.
When to widen or tighten
Change one thing at a time, then give it two nights.
| What you see | What it means | Change this |
|---|---|---|
| Under 2 Strong for two nights | The pool is too small or the brief too strict. | Widen in this order: add adjacent titles, widen the distance, move one must-have to "ideally". Never drop the location for an onsite role. |
| More than half dropped | The search is finding the wrong job under the right title. | Read the drop reasons. Add the wrong job's words to "Leave out", and one word only the right job uses to must-haves. |
| The same companies every night | You've emptied the obvious ring. | Add companies in a different industry that need the same skill, and companies their people left. |
| Strong people, no replies | It's the message or the timing, not the search. | Keep the search. Fix line 1 of the drafts and lead with people who've been 2 to 4 years in their role. |
| Nothing new for a week | The market for this role in this place is small. | Tell the hiring manager with numbers. Discuss remote, relocation or the level. |
What it costs
No surprises.
Set it up tonight. Read the list tomorrow.
Create your free The Cognitive account, connect it to Claude, and paste the prompt with one search in it. 7 days free, 100 search credits, no card.