Writing Cold Outreach

Goal

Produce a ready-to-send cold email and LinkedIn connection request for each lead, where the only AI-written words are the one line that proves this wasn't a mass send.

The rule that matters

Never let the model write the whole email. The template is fixed — the user wrote it, tested it, and it works. The model fills exactly one slot: the observation about this company.

Hi {first_name},

{opener}                         <-- the ONLY AI-written line

Most companies scaling this fast hit the same bottleneck: {pain_point}.

We helped {similar_company} solve this and saw {metric} in 60 days.

Worth a 15-min chat to see if we can do the same for {company}?

{sender_name}

Everything except {opener} comes from CLI arguments or the lead record. Same words every time. If the user asks you to "make the emails more varied" or "let the AI write the whole thing", explain why that's the wrong direction: variance in the body is what makes a sequence untestable and what makes replies stop.

To change the email itself, edit templates/email_template.md. Not the code.

Prerequisites

Leads must have been through research_lead.py and have fact1..fact5 populated. Leads without facts are skipped and flagged, not given a generic opener.

Process

Step 1 — dry run, always

python -X utf8 scripts/04_outreach/write_outreach.py \
    --input output/researched.json --dry-run \
    --sender-name "Armin" \
    --pain "leads going cold because nobody follows up past email one" \
    --similar-company "a 12-truck HVAC company in Mobile" \
    --metric "31 extra booked jobs in 60 days"

Prints one complete email. Read it out loud. If the copy is wrong, fix the four arguments — those words go into every single email.

Step 2 — the full run

python -X utf8 scripts/04_outreach/write_outreach.py \
    --input output/researched.json --out-csv output/outreach.csv \
    --sender-name "Armin" --pain "..." --similar-company "..." --metric "..."

What gets enforced in code

The prompt asks for these. The script guarantees them, with one retry that feeds the specific failure back:

• Subject starts with Quick question about {Company}'s, max 11 words
• Body starts with I noticed {Company} recently, max 15 words
• No "Congratulations", "Congrats", "amazing", "great work", "impressive"
• No years, no month names, no quoted event names
• LinkedIn request max 200 characters, and none of "building my network" / "expanding my network" / "also working in"

Anything still failing after the retry is held back with a reason in outreach_status, not shipped.

Reading the output

Check outreach_status before sending anything:

Value	Meaning
ok	Passed every check
skipped: no research facts	Nothing true to say — go research it or drop it
email failed validation: ...	Model couldn't hit the constraint twice; read the reason

Anti-patterns

• Sending the whole list without reading five of them first
• Writing an opener for a lead with no facts (that's what the flag is for)
• Putting the pain point or social proof into the AI prompt — they're your copy, they're arguments
• Editing the validators to make more leads pass. If leads are failing, the research is thin.

Follow-up

templates/email_5_touch_sequence.md has the cadence. The course's own number: roughly half of people stop after one message, and most deals land somewhere between touch three and five.

Cost

About $0.005 per lead.