Drop-off Rescue

Push stalled funnel contacts to the next step. This skill finds people who signed up but didn't convert, drafts stage-appropriate re-engagement emails, and (with an ESP MCP connected) schedules the send after explicit confirmation. Re-runnable weekly without double-sending — the skill tracks what was already sent per stage and skips contacts still inside their cooldown window.

Works in three modes:

• cloud — Cogny MCP + ESP MCP: full automation, cross-ESP dedupe, findings
• solo — ESP MCP only: tag-based dedupe, per-ESP automation
• free — no MCP: pasted CSV of stalled contacts, draft-only output (you send manually)

Usage

/drop-off-rescue              # weekly rerun — default
/drop-off-rescue weekly       # same
/drop-off-rescue audit        # read-only: report stalls + what would be sent, write nothing
/drop-off-rescue free         # no MCP — paste CSV of stalled contacts

Designed to be scheduled weekly via /loop or /schedule. On reruns, already-nudged contacts are skipped; newly-stalled contacts are picked up; contacts that progressed get celebrated in the output.

Step 1 — Mode detection

Probe both MCP namespaces:

1. If mcp__cogny__<svc>__* tools are present for at least one ESP → cloud mode (and record that ESP as connected).
2. Else if mcp__<svc>__* tools are present for at least one ESP → solo mode.
3. Else → free mode.

If the argument is free, force free mode regardless of what's connected.

If multiple ESPs are connected, ask the user which one to run against (or offer to run all and cross-dedupe in cloud mode).

Step 2 — Canonical funnel stages

The skill works against a canonical stage taxonomy. The user can extend it, but these slugs drive tag names, campaign naming, and dedupe state — keep them stable across runs.

E-commerce:

SaaS / trial funnel:

Cogny's own Solo funnel (signup → no-channel → no-trial → no-client → no-subscription) is a good reference implementation of this taxonomy.

If the user describes a funnel with stages that don't map to these slugs, add new kebab-case slugs — just keep them stable across reruns (the slug is the dedupe key).

Step 3 — Resolve current stage per contact

Primary signal: ESP tags / segments. Read contacts tagged stage:<slug> using the per-ESP tool:

Klaviyo event-derived fallback (only when stage:cart-abandoned tag is absent): query list_events with metric="Added to Cart" in the last 7 days, exclude profiles who have a subsequent Placed Order. Write the stage:cart-abandoned tag back via update_profile so the next run is consistent.

Conversational fallback (any ESP, when no stage tags exist at all):

1. Ask the user: "I don't see any stage:* tags on your contacts. Which of these funnel stages apply to your business?" — list the canonical slugs.
2. For each picked stage, ask: "Who is on this stage? Describe the condition — e.g. joined list > 3 days ago AND no purchase, or clicked email X but never visited /signup."
3. Translate each heuristic into the ESP's available filters (list membership, tags, date ranges, clicked links) and preview the matching count.
4. Offer (under the confirmation gate in Step 7) to apply stage:<slug> tags now, so future runs are fully automated.

Step 4 — Already-sent detection (the core mechanism)

Weekly reruns must not re-send the same stage email to the same contact. The skill uses a three-layer hybrid:

Layer 1 — contact tag (primary, all modes with an ESP connected)

On every successful send, write three tags to the recipient:

• rescue-sent:<slug>:<YYYY>-W<WW> — append-only; one per week nudged
• rescue-last-stage:<slug> — overwritten; always reflects the latest stage nudged (enables stage-advancement detection in one read)
• rescue-last-sent:<YYYY-MM-DD> — overwritten; used for cooldown math and the global 7-day anti-fatigue rule

Why tag-based primary: works in solo mode without a context tree, survives cross-machine usage, all four ESPs expose a tag-write tool, debuggable by a human in the ESP UI, bulk-clearable if something goes wrong.

Layer 2 — ESP campaign-name convention (fallback)

Every rescue campaign is named exactly:

drop-off-rescue / <stage-slug> / <YYYY>-W<WW> / <variant A|B>

On rerun, if a contact's tags were stripped (manual edit, sync tool), list_campaigns / tool_list_campaigns / tool_list_sent in the last 35 days is scanned for matching names. Recipients are cross-referenced via the per-ESP report tool (list_events, tool_get_report, tool_get_campaign_statistics, tool_get_report).

Deliberately no markers in subject or preheader — the user-visible inbox copy stays clean.

Layer 3 — Cogny context-tree log (cloud mode only)

Write on every send:

insights/drop-off-rescue/<slug>/<sha256(email)[:12]>/last-sent
  → { "week": "2026-W17", "variant": "A", "esp": "klaviyo", "ts": "2026-04-24T10:00Z" }

Read on the next run. This layer catches cross-ESP duplication — the same contact living in both Klaviyo and Mailchimp under the same org won't get two nudges in the same week.

Dedupe decision

A contact is classed "already nudged at current stage within cooldown" if any of the three layers reports a hit inside the cooldown window. Cooldown defaults (printed in output, user-overridable):

Step 5 — Classify every contact (the weekly loop)

For each stage, read the full stage:<slug> cohort, then classify each contact: