email-generation

Email Generation

Generate cold outreach emails from a contact CSV + prompt template. The prompt template is self-contained — it has all voice, research, value prop, proof points, and personalization rules baked in. This skill just runs it per row.

Architectural Principle

This skill is a runner, not a reasoner. All strategic reasoning (voice, value angles, proof points, research data) was done by the email-prompt-building skill at prompt-build time and embedded in the prompt template. This skill reads the prompt + CSV and generates emails. It does NOT read the context file, hypothesis set, or research files.

prompt template (.md) ─┐
                       ├──▶ generate email per row ──▶ emails CSV
contact CSV ───────────┘

Inputs Required

Input	Source	Required
Contact CSV	File with recipient data + enrichment columns	yes
Prompt template	.md file from email-prompt-building skill	yes

That's it. No context file, no hypothesis set, no research files.

Contact CSV Columns

The prompt template specifies which columns it needs. Check the prompt's "Enrichment data fields" section for the expected column names. Common columns:

Required (always):

• first_name, last_name, company_name, job_title

Enrichment (campaign-specific): Listed in the prompt template. If the prompt references a field that's not in the CSV, the email quality degrades. Check column alignment before running.

Name Sanitization

Before generating emails, run scripts/sanitize-names.py on the contact CSV:

python3 scripts/sanitize-names.py <contact.csv> [output.csv]

The script strips titles (Dr, Prof, etc.), removes rows with single-character names, emoji, junk values (N/A, Test, -), and fixes all-caps casing. It outputs a *_sanitized.csv and prints what was cleaned/removed.

Review the removed rows before proceeding. Do not generate emails for rows with invalid names.

Running the Generator

Script-first, not in-context. Always generate via a script that calls the API per contact. Never generate emails inside the conversation — it's slow, expensive, and impossible to rerun after prompt edits.

Step 1: Dry run

Before spending API credits, show the user a dry run:

1. Read the prompt template and contact CSV
2. For 2-3 sample contacts, display exactly what data will be passed (all enrichment fields, hypothesis match, structural variant selection)
3. Ask the user to confirm the data looks correct before proceeding
4. If enrichment fields are missing or misaligned, flag it and stop

Step 2: Generate via script

Write a generation script that reads the prompt template + contact CSV, calls the API per row, and writes output files. See references/generation-script.md for the script template and implementation details.

Adapt the script to the user's API setup (Anthropic, OpenAI, etc.) and the specific prompt format.

Step 3: Output both CSV and MD

Always generate two output files:

• claude-code-gtm/csv/output/{campaign-slug}/emails.csv — for upload to sequencer
• claude-code-gtm/csv/output/{campaign-slug}/emails.md — for human review (one email per section, with contact name and company as headers)

Quality Checks

After generating, verify:

• Every email is within the word limit specified in the prompt
• No banned phrases from the prompt template appear
• Enrichment data was actually used — not just generic text
• Example queries in P2 are specific to each recipient's verticals
• Proof points vary across emails (not the same PS for everyone)
• Subject lines meet the prompt's length constraints

Segmentation-Aware Generation

When the contact CSV includes segmentation data (from list-segmentation):

Tier 1 companies:

• Generate individually with full attention to enrichment data
• Route through email-response-simulation for review before sending

Tier 2 companies:

• Group by hypothesis_number
• Generate in batches within each hypothesis group
• Spot-check 2-3 from each group

Tier 3 companies:

• Do not generate emails
• Route back to list-enrichment or list-building

Feedback Loop

When the user gives feedback on generated emails, the workflow is always:

1. User identifies what's wrong (tone, structure, missing data, wrong angle)
2. Update the prompt template — the fix must be systemic, never a one-off edit
3. Rerun the script with the updated prompt
4. Review the new output

Never hand-edit individual emails. If one email is bad, the prompt is bad — fix the source. Track changes made to the prompt so the user can see the evolution.

Building a New Prompt Template

If no prompt template exists for this campaign, use the email-prompt-building skill to build one. That skill reads the context file and research, then synthesizes a self-contained prompt. Do not build prompts ad hoc in this skill.