Cold email

BeReach finds the work email; Claude writes the drafts. The user sends them from their own mailbox or email tool. Nothing here sends or reads an email. bereach_draft_dm, bereach_send_dm and bereach_schedule_drafts are for LinkedIn messages: never use them for an email.

Each step ends with a Deliver: a page drawn as in references/artifacts.md. Only a page that spends email credits or hands over emails to send waits for the user's yes; the Sending check is shown and Claude goes on.

Refused with a trial link: give it to the user and stop, no retry.

The web only teaches the user's trade; a vendor statistic never decides. Asked for best practice: a verdict per slot, then the upgraded email to validate.

Step 1. Sending check (once per user)

Five questions only:

1. Which mailbox sends (work domain, not a free webmail)?
2. Are SPF, DKIM and DMARC set? If unsure, say how to check (references/deliverability.md).
3. How many cold emails a day? Over 50 per mailbox: too many.
4. Which countries are the people in (sets the footer, references/legal.md)? Germany: left out unless the user says otherwise. An EU country legal.md does not name: amber, check.
5. Is tracking off (no open pixel, no wrapped links)?

Deliver: "Sending check".

Step 2. Who, and why them

1. Read the list: bereach_contacts with the list's campaignId, engaged: false (a found whyNow is a free dated reason). Drop do-not-contact people and anyone tagged cold-email, silently (count only).
2. Read what they sell: bereach_memory_list, type offering, fullContent: true. Its ## Personas holds the persona cards (each with Offer and Why now window).
3. Read the voice card: bereach_memory_list, type playbook, fullContent: true, the one place for writing rules (it wins on voice, the self-check on craft).
4. Persona cards or voice card missing: build them first (linkedin-personas or linkedin-voice skill).
5. Before sorting, read linkedin-personas references/persona-card.md (ICP, plan groups and cards). Put each person in one persona, tag persona:<short-name> (bereach_contact_update, tags.add). READY: title matches one card, company inside its ICP. CHECK: two cards fit or the company is outside; no email lookup, brief or draft until the user decides. Out of scope (not this persona, outside the ICP): left at the search step; found later: HOLD, no email. Batches of 25 at most, 2 per company.

Deliver: "Who is who", cost for READY people only, plus the whole-list signals ceiling (bereach_add_signals start: false). Nothing is spent before the yes.

Step 3. One sequence per persona

1. Open references/frameworks.md now, before writing a word (Default, Which shape when). Baseline: its default (self-check, opening and ask) in the user's voice unless a row fits better, opening on a dated fact from their brief, else a moment of their work. The user's own emails give the voice (address form, greeting, sign-off, words), never the slots: keep a slot of theirs only where it passes the self-check, rewrite every failing one with a framework named on the page, replies included. The offer comes from the user's own trade, is kept by the reader (a sheet, a plan, a sample of their work; software: a free audit or estimate of THEIR case, never a case study), exists before the send (no blank); an object a past reply asked for works as the offer. Never offer to find contacts unless that is the user's trade.
2. 3 or 4 emails (references/sequence.md), days 1, 4, 8, 14, step 2 a reply in the thread, each ending on one yes/no question naming its object, the last offering the way out. ONE offer per sequence.
3. Proof only in a follow-up, from a real customer of the user, else none. Never a claim about the user's work, customers or results they did not give: ask.
4. A {why now} slot in every template that reads right whatever fills it. No fact under 6 months: a moment line. Two languages: same offer, same ask.

Deliver: "Message templates", one per persona, email 1 filled for one real person. After the yes, save it under that persona: bereach_memory_save type offering replaces the whole entry, so resend the Step 2 offering whole.

Step 4. Emails, then the reasons

1. bereach_find_emails with every READY contactId in ONE call (25 at most). One email credit per email found; none found is free. Quote emailCreditsUsed from the reply, never a sum.
2. Only verified: true is sendable; else "unverified, skip".
3. Email credits run out: give buyCreditsUrl and the note, stop for the people left.
4. bereach_brief (contactId, campaignId, no include, historyVerbatim 200; maxAgeDays left at its default, the conversation is read anew) for every READY person before each email to them, batches included. For a whole list, bereach_add_signals, within the Who is who yes. Nothing usable: a moment line, marked.
5. Fill {why now}: one sentence giving the fact AND why it made the user write (the tie to the offer), tense per its date (references/frameworks.md, Personal line rules). Never invent a fact.

Step 5. Drafts and review

Open references/frameworks.md again, then write each READY person's email 1 from the approved sequence plus their line. Self-check and fix every draft. An offer not deliverable now: block those rows.

Deliver: "Drafts for review". The user approves, edits or removes rows.

Step 6. Hand over

Deliver: "Ready to send", then:

• bereach_contact_update on everyone handed over: tags.add ["cold-email", "cold-email-YYYY-MM-DD"]. Never tags.set; leave notes alone.
• One line: own mailbox, 20 to 50 a day at most, tracking off, found addresses not emailed deleted (references/deliverability.md, legal.md).

Step 7. Replies and results

• Stop, unsubscribe, not interested: bereach_contact_update with doNotContact: true; the user removes them from any running sequence the same day.
• A positive reply (pasted by the user): bereach_brief first (maxAgeDays left at its default, the conversation is read anew); answer what they asked with one next step, the same working day; a fact only the user knows is a blank, never invented.
• Results: the user's own counts (mailbox, tool, CSV, pasted replies), never estimated. A rate counts only batches 7 days past their first email; younger ones show "counting". No rate under 10 sent. One change per batch between two passing versions, within one persona. While a test runs, send the version with more interested replies; a tie keeps the baseline.

Deliver: "Results".

Self-check (every draft)

• No title, tenure, headline or company description recited; the fact is tied to why the user writes.
• Fact in their row or brief, tense matching its date, under 6 months. Nothing invented, no figure without a source, no guess in quotation marks.
• Line one names their situation, not us; then who writes (name + trade noun, one clause); one offer from the user's trade, no proof in email 1, one yes/no question naming it. No call, slot, meeting or vague wish.
• No hedge, no guessed feeling, no thank-you or line that says nothing about them.
• Email 1 body 50 to 90 words, follow-ups under 50 (footer not counted), personal line 240 characters at most.
• Subject 1 to 5 words, no fake "Re:"/"Fwd:"; no buzzword; no link, image, attachment or tracking in email 1.
• Native-grade language per the voice card. French: vous unless the card says tu or the person wrote tu first; "Bonjour" capitalised with vous.
• Country footer on email 1 (references/legal.md); follow-ups carry the opt-out line only (US: full footer on every email).
• 4 emails at most per person; never the same email to one company in the same week; no 3 drafts share opening or shape (rotate the moment lines).

What not to do

• Promise reply rates or quote open rates.
• Send to a guessed or unverified address.
• Write to anyone who asked to stop, on any channel, or anyone on HOLD.