cold-sweep

name	cold-sweep
description	Drain overdue CRM "red debt" into the Action Queue as routed, voice-drafted nudges for one-click CEO approval. Use when the CEO says "cold sweep", "drain the cold contacts", "sweep the overdue CRM", or wants the overdue-contact backlog turned into ready-to-send drafts. NOT for sending email (that is the human-approved executor) - this skill only drafts into the queue. NOT for a single follow-up (use /follow-up). NOT for pipeline review (use /crm). CEO-only (not synced to executives) during the prove-out.
argument-hint	[--dry-run]
allowed-tools	Read, Write, Bash(python3:*), Bash(python:*), Skill
metadata	{"author":"Misha Hanin","email":"[e-mail]","version":"1.0"}
x-heading-orchestration	{"parallel_safe":"partial","shared_state":["outputs/operations/action-queue/"],"triggers":["cold sweep","cold-sweep","drain cold contacts","sweep overdue contacts","drain the red debt"]}
x-heading-capability	{"what":"Turns the overdue-contact \"red debt\" from crm-health into ~150-word voice-drafted nudges and deposits them in the Action Queue as email_send cards marked ready_for_review for one-click CEO approval.\n","how":"Run /cold-sweep (or /cold-sweep --dry-run to preview routing only). It fills needs_draft cards via scripts/action-queue.py through the daemon; the CEO then approves on the bridge Action Queue page or with scripts/action-queue.py approve <id>. CEO-only, not synced to executives.\n","when":"Use to drain the whole overdue-CRM backlog into ready drafts. For a single follow-up use /follow-up; for pipeline review use /crm. It never sends - the human-approved executor does that.\n"}
x-heading-routing	{"category":"Operations","triggers":["cold sweep","cold-sweep","drain cold contacts","sweep overdue contacts","drain the red debt"],"exclusions":["Single follow-up -> /follow-up","pipeline review -> /crm","sending -> human-approved executor (never this skill). CEO-only, not synced to executives."],"compound":"No","router":"auto"}

/cold-sweep

Turn the overdue-contact backlog into prioritised, cited, voice-drafted nudges sitting in the Action Queue for one-click go/no-go. The deterministic routing (who, what priority) is done by scripts/cold_sweep_core.py; this skill does only the part that needs judgment and voice - writing the actual draft body.

Split of responsibilities (plan 2026-06-03, Design Decision 5):

• Deterministic + headless: cold_sweep_core reads crm-health.py --json, routes each overdue contact, and deposits email_send cards with draft_status: needs_draft. This runs either on the daemon's schedule (06:30 local time, when enabled) or via scripts/cold-sweep.py for a manual run (daemon-free, in-process append_cards).
• Voice + judgment (this skill): fills the needs_draft cards with a ~150-word nudge in Misha's voice and flips them to ready_for_review.

The CEO then approves+sends from the terminal with /queue approve <id> (or scripts/action-queue.py approve <id>) - a SYNCHRONOUS, watched send, daemon-free. This skill never sends; it only drafts into the queue.

Phase 0 - Load context

1. Read reference/misha-voice.md (voice, maritime inventory, "what Misha never says").
2. Apply .claude/rules/voice.md, .claude/rules/humanization.md, .claude/rules/hidden-chars.md, .claude/rules/voss.md.
3. Read the queue store outputs/operations/action-queue/queue.json (read-only) to discover cards. The targets are cards with action_type: "email_send" AND status: "pending" AND draft_status: "needs_draft".

Phase 1 - Ensure cards exist

If there are no needs_draft cards in the queue:

• Run a manual sweep: python3 scripts/cold-sweep.py (deposits DAEMON-FREE via the in-process append_cards; works with the bridge daemon down).
• --dry-run (if the CEO passed it): run python3 scripts/cold-sweep.py --dry-run and present the routed cards without drafting. Stop after presenting.

Re-read queue.json after a manual sweep to pick up the new cards.

Phase 2 - Draft each needs_draft card (in voice)

For each needs_draft email_send card (cap at the cards present; do not invent contacts):

1. Read the linked contact file at card.contact_file (e.g. crm/contacts/<slug>.md) for the relationship, last interaction, and any open commitments. Use the card's citations for the cadence-breach context.
2. Draft a ~150-word nudge in Misha's voice: direct opening, one concrete specific from the contact's history, one clear ask. Hyphens, never em-dashes. No banned vocabulary. Do not fabricate facts - if the contact file lacks a hook, keep it short and honest rather than inventing one.
3. Write the body to a temp file (e.g. outputs/documents/_work/cold-sweep-<id>.txt).
4. Validate the draft: python scripts/sanitize-text.py <tmp> --scan and python scripts/humanization-check.py <tmp>. Fix any findings before finalizing.
5. Finalize the card through the daemon (single-writer; flips draft_status to ready_for_review): python3 scripts/action-queue.py edit <id> --subject "<subject>" --body-file <tmp>

Draft cards sequentially (the queue is shared state; the daemon serialises writes, but sequential drafting keeps the run legible).

Phase 3 - Report

Summarise: N cards drafted and flipped to ready_for_review, M skipped (and why). Remind the CEO:

• Review with /queue (or python3 scripts/action-queue.py list); approve+send with /queue approve <id> (or python3 scripts/action-queue.py approve <id>).
• approve SENDS synchronously and is watched (no async executor). Dismiss with python3 scripts/action-queue.py dismiss <id> suppresses re-proposal for 14 days.

Then run /brain-audit --sources crm/contacts --entity cold-sweep if a synthesized summary was produced; append the footer. (Cold-Sweep is mechanical routing + per-card drafting, so the audit is optional here.)

Voice rules

• Misha's voice: direct, specific, committed. Hyphens, never em-dashes (-- and — both banned in drafts). Tribe vocabulary, ODUN.ONE, DPI+.
• Every draft carries at least one named/dated specific from the contact's real history. If none exists, keep it short - do not fabricate.

NEVER

• NEVER send email. This skill drafts into the queue only; the human approves and the executor sends.
• NEVER write queue.json directly. All card mutations go through scripts/action-queue.py (the daemon is the single writer).
• NEVER invent contacts, interactions, or specifics not present in the contact file.
• NEVER fabricate a hook to hit ~150 words - a shorter honest draft beats a padded one.
• NEVER mark a card sent or approved on the CEO's behalf.