CLEANSKIN — kill the AI dash

The em dash (—) and its shorter cousin the en dash (–) are the single most recognisable "this was written by AI" signal. A normal keyboard hyphen is -. CLEANSKIN removes em/en dashes everywhere and swaps in punctuation a human would actually type.

The rule (never violate)

Never emit an em dash (—) or en dash (–) in any copy you write for Will — cold emails, captions, ads, headlines, landing pages, Klaviyo, anything. Replace it with:

• a normal hyphen - for "topic - name" subject/headline separators and number ranges (2-3),
• a comma for a clause aside ("across Brisbane, steady inbound"),
• a full stop when it joins two sentences,
• or nothing (drop it — e.g. the sign-off — Will becomes just Will).

This applies to copy you generate AND to copy you clean for him.

The tool — scrub.py

Deterministic, regex-based, reversible (writes .emdash.bak before editing files). Three modes:

# 1. Clean a string (or piped text)
python3 scrub.py text "we built voice agents — they answer after hours"
echo "$DRAFT" | python3 scrub.py text

# 2. Clean file(s) in place (.bak saved per file)
python3 scrub.py file path/to/copy.md path/to/prompt.yaml

# 3. Clean a LIVE Instantly campaign — per-lead subject/body (custom_variables +
#    personalization) AND the follow-up sequence-step templates.
python3 scrub.py instantly <campaign_id> --dry   # preview counts, no writes
python3 scrub.py instantly <campaign_id>         # apply (idempotent; re-run safe)

Instantly mode reads INSTANTLY_API_KEY from ~/Desktop/ADMIN/.env, leaves real business names (companyName/business_name) untouched, treats email_1_subject/personalization as subjects (hyphen) and everything else as prose (comma), and retries on rate-limits. Back up first for large live campaigns — dump leads to JSON before a multi-thousand-lead run.

Where it's wired in already

The cold-email generator (~/.claude/skills/cold-email-ai-agents/run.py) now scrubs every generated subject + body through this same logic before it ever reaches Instantly, and its validator rejects any em/en dash. So new batches are clean at the source; CLEANSKIN is for cleaning existing copy, other pipelines, and one-off text.