The AI Sourcing Setup

STEP-BY-STEP N8N WORKFLOW SETUP

1. Create a new n8n workflow and add a Schedule Trigger node set to run nightly (e.g. 11pm).
2. Add an HTTP Request or scraping node (an Apify actor works well for LinkedIn/Indeed) that pulls new profiles matching your saved search criteria.
3. Add a Split In Batches node to process profiles in groups of 20-50, avoiding timeouts on large runs.
4. Add a Claude node inside the batch loop that scores each profile against your brief.
5. Add an IF node that filters for scores above your threshold (7/10 is a reasonable starting point).
6. Add a Google Sheets append node for qualified candidates, and a Telegram node for the morning summary.

CLAUDE API CONFIGURATION

1. Generate an API key at console.anthropic.com and add it as a credential in n8n.
2. Write a system prompt that includes your full brief and a strict output format (score, one-line reason, green/red flags) so the output parses cleanly into the sheet.
3. Set a low temperature (around 0.2-0.3) for scoring tasks so the model is consistent across hundreds of profiles rather than creatively varying its judgment.
4. Budget for cost: at this volume, expect roughly £20-40/month depending on model choice and profile volume.

TELEGRAM BOT SETUP

1. Message @BotFather on Telegram, run /newbot, and save the API token it gives you.
2. Add the token as a credential in n8n and connect a Telegram node at the end of your workflow.
3. Format the morning message as a short summary: total profiles processed, number qualified, and a link to the sheet. Keep it under 5 lines so it's genuinely a 30-second read.

GOOGLE SHEETS TEMPLATE

Columns: Name | LinkedIn URL | Current Role | Current Company | Score | Reasoning | Outreach Status | Date Added. Add a filter view for "Outreach Status = Not Started" so your daily review only shows what actually needs action, not the full historical log.

TROUBLESHOOTING GUIDE

• Agent surfaces too few candidates: loosen the Boolean search criteria or lower the score threshold slightly, one variable at a time so you can see what moved the needle.
• Agent surfaces too many low-quality candidates: tighten the brief's non-negotiables. A vague brief produces a vague score spread.
• Workflow fails silently overnight: add an error-handling branch that sends you a Telegram alert if any node fails, so you catch it the next morning instead of days later.
• Costs creeping up: check whether you're re-processing the same profiles repeatedly. Add a dedup check against existing sheet rows before scoring.

THE GUARDRAILS THAT KEEP QUALITY HIGH

Review a random 10% sample of scored profiles by hand each week, even once the system is stable. This is the single habit that catches drift early, before a subtly wrong brief interpretation quietly lowers pipeline quality for weeks. Re-run the brief itself through the Role Killer or JD Writer prompts every quarter, since roles and priorities shift faster than most people update their sourcing criteria.