Email Drafting

Email Drafting

Produce cold-email copy that follows proven frameworks and hard quality rules. The agent does all the writing (this is a reasoning skill); the bundled email_lint.py only verifies the output against the hard rules deterministically.

When to use

• "Write a cold email" / "draft outreach" / "write a sequence" / "help with email copy."
• Any task that needs cold-email copy — including the signal composites (champion-move / funding / hiring / leadership / news) and cold-email-outreach, which call this skill for the drafting step.

How to draft (agent steps)

1. Intake campaign context (product, problem, angle, trigger/signal), audience (title/seniority/dept, frustrations, objections), proof (peers, metrics — optional), tone, sender, sequence config, personalization tier. Ask everything at once; skip answered.
2. Pick the framework by angle:
   • PAS (Problem-Agitate-Solve) — pain-signal driven.
   • BAB (Before-After-Bridge) — aspirational/outcome driven.
   • AIDA (Attention-Interest-Desire-Action) — cold-database, no specific signal.
   • Signal-Proof-Ask — when a concrete trigger/signal exists (the composites' default).
3. Pick the personalization tier by lead volume vs ROI: 1 generic, 2 segment, 3 deep.
4. Draft each touch to word-count targets — opener (T1) 50-90, follow-ups 30-50, breakup 20-40 — structured Hook -> Evidence -> Offer.
5. Write subject lines from the proven patterns: < 50 chars, lowercase-ish, no emoji, no ALL-CAPS, never "quick question". Curiosity / specific-benefit / relevant-question.
6. Design cadence — T1 day 1, T2 day 3-5 (new angle, not a nag), breakup last.
7. Enforce hard rules — never lie about how you found them, never fabricate metrics or peer companies, exactly one CTA per email.

How to run (the linter)

After drafting, write the draft(s) as JSON and verify them. The linter is Python 3 stdlib only — no install:

python3 ${SKILL_DIR}/scripts/email_lint.py --input ${WORKSPACE}/sequence.json

JSON shape — one object or a list:

[
  {"touch_number":1,"send_day":1,"touch_type":"opener","framework":"Signal-Proof-Ask",
   "subject":"...","body":"..."},
  {"touch_number":2,"send_day":4,"touch_type":"followup","subject":"...","body":"..."},
  {"touch_number":3,"send_day":9,"touch_type":"breakup","subject":"...","body":"..."}
]

touch_type is one of opener | followup | breakup (drives the word-count target).

The linter exits 1 on any ERROR (over-length subject, banned phrase, emoji in subject). Regenerate the offending touch — never truncate to fit. Warnings (word count off target, 0 or >1 CTA) are advisory.

Outputs

• The drafted email(s)/sequence as the result: per touch {subject, body, word_count, framework, touch_number, send_day} plus the chosen tier/tone — returned for review or persisted to a workspace draft / channel attachment for an orchestrator (cold-email-outreach, signal composites) to load into a sending tool.

Credentials / env

• env.required: none for the linter. The drafting itself is done by the host agent (the reasoning model) — no LLM API key is consumed by this skill.
• env.optional: none. No paid service is applicable — this is a pure drafting + lint skill with no external-data step, so the paid-services-optional policy adds nothing here.

Notes & edge cases

• This is a copy generator, not a sender — hand off to cold-email-outreach / lemlist / instantly for the actual send.
• If the linter flags an over-length subject, regenerate (rewrite shorter), never truncate.
• Never fabricate metrics or peer companies; if proof is absent, fall back to a softer curiosity/question angle rather than inventing claims.
• The champion-move composite intentionally flips the "never say how you found them" rule (you reference a real prior relationship) — that is the one allowed exception.