Reply Triage

Reply Triage

Outreach sequences come back with replies, and a brand staring at a stack of "interested", "has a question", and "declined" responses needs to know what to actually DO with each one. This skill pulls the brand's replied connections, reads each reply's classification + brand-safe intent summary + confidence, and sorts them into three action buckets — INTERESTED (the matchmaker is already brokering the intro; nothing to do), QUESTION (the creator wants info first; the skill DRAFTS a short brand-safe answer to send), and DECLINED / opted-out (acknowledge, optionally swap in an alternate from the session's shortlist). The deliverable is a prioritized triage memo with the drafted answers inline. It is the next-steps layer; for a board/status view of the whole pipeline use connection-pipeline-tracker. Free/cheap — it only reads, never reaches.

Read first: ${CLAUDE_PLUGIN_ROOT}/shared/conventions.md (tool schemas, credit prices, the nine conventions) and ${CLAUDE_PLUGIN_ROOT}/shared/connection-flow.md (the connection contract, entitlement/degradation, and the privacy invariants — this is a connection-enabled skill). This skill honors thrifty/thorough credit modes (${CLAUDE_PLUGIN_ROOT}/shared/credit-modes.md).

This runs AFTER outreach has gone out and replies have landed (e.g. from cast-and-connect). It is read-mostly: it calls only list_connections and get_connection_status (both free) — it does NOT call request_creator_connection. Drafting an answer never reaches anyone; the brand sends what they approve and the matchmaker relays it.

Inputs to collect

• Focus (optional) — a campaign type and/or a status to narrow to (e.g. "just the questions", "the SS26 campaign"). Default: all replied statuses, newest-first. Ask only if the user implied a filter you can't infer.
• Alternates source (optional) — if a shortlist from an earlier in-session run exists, reuse it to suggest a swap for any DECLINED creator. Never go and build one from scratch here — that's brief-to-shortlist's job; if none is in session, the DECLINED bucket just acknowledges without a swap.
• Credit mode — default thorough; thrifty on the standard triggers. Both modes are free or near-free here (see Credit footprint).

Flow

Step 0 — Entitlement / graceful degradation. The connection tools are entitlement-gated (creator_connections, pro + pilot). If list_connections returns the refused envelope (plan without connections), do NOT error: say "Connections aren't on your plan, so there are no outreach replies to triage — here's how to enable outreach," and stop cleanly. Same graceful path if the list is simply empty: "No replies yet — your sequences are still in flight (or no outreach has gone out). Nothing to triage."

Step 1 — Pull replied connections (free). Call list_connections once per replied status you're triaging (the API filters on a single status):

list_connections { "status": "replied_interested", "limit": 100 }
list_connections { "status": "replied_question",    "limit": 100 }
list_connections { "status": "replied_declined",    "limit": 100 }
list_connections { "status": "opted_out",           "limit": 100 }

(Honor a narrower focus if the user gave one — e.g. only replied_question.) Each row carries conn_ref, status, campaign type, next scheduled touch — and no contact info ever. 0 credits.

Step 2 — Read each reply's detail (free). For every replied connection, call get_connection_status to get the classification + the brand-safe intent summary + confidence + touches sent:

get_connection_status { "conn_ref": "<ref>" }

0 credits each. Also read each row's engagement block (connection-flow.md) — whether the creator clicked or opened before replying is useful context for the draft (a clicker has already looked at what you linked; a click is reliable intent, an open directional only, and tracking_enabled: false means the send predates tracking — unmeasured, not disengaged). Work ONLY from the classification + the brand-safe intent summary the server returns — the server already PII-scrubs reply content, and this skill must never surface raw reply text (which could carry an address or handle the creator typed). If a returned summary still looks like it contains contact details, drop it to the bare classification rather than quoting it.

Step 3 — Bucket + draft (no tool calls).

• INTERESTED → the matchmaker is already brokering the double-opt-in joint intro. Action = nothing to do but wait; tell the user what to expect (an intro once the creator confirms; contact is still never exposed by these tools). Never promise it converts to a booking.
• QUESTION → the creator wants info before deciding. Draft a short, brand-safe reply addressing the intent summary (scope, timeline, comp band, usage, exclusivity — whatever the summary flags). The draft is a suggestion the brand sends; the matchmaker relays it. It handles no contact info and invents no quotes — fill comp/budget from a corpus rate band only if the brand already supplied one, else leave a [brand to confirm] placeholder.
• DECLINED / opted-out → acknowledge (no re-reach — opted-out creators are suppressed and reaching them is harassment risk). If an in-session shortlist exists, suggest one not-yet-contacted alternate per declined slot; otherwise just note the slot is open.

Step 4 — Prioritize. Order the memo QUESTION-first (they're the time- sensitive, action-required set), then INTERESTED (FYI), then DECLINED (cleanup).

Deliverable

A prioritized triage memo in markdown:

# Reply Triage — <campaign / "all outreach">
_Creatorland Data · inbound replies as of <date>_

## Counts
- ✅ Interested: <N>   (matchmaker brokering intro — nothing to do)
- ❓ Question: <Q>     (needs a reply — drafts below)
- ⛔ Declined / opted out: <D>

## ❓ Questions — drafts to send  (do these first)
For each:
**<campaign type> · conn <ref-short> · confidence <hi/med/lo>**
- They're asking: <brand-safe intent summary — no raw reply text, no contact>
- Engagement before replying: <"clicked <when>" / "opened <when> (directional)"
  / "no signal" / "sent before tracking — unmeasured">
- Suggested reply (you send; matchmaker relays):
  > <2–4 sentence brand-safe draft answering the question; comp as a corpus
  > band or [brand to confirm]; no contact handling; no invented quotes>

## ✅ Interested — what to expect
<one line per interested conn: "conn <ref> (<campaign>) — matchmaker is setting
up the double-opt-in intro; you'll get a joint intro once they confirm. Nothing
to do.">

## ⛔ Declined / opted out
<one line per: acknowledge; if a session shortlist exists, "suggest swapping in
@<alt> (not yet contacted)"; else "slot open — re-cast if you want a backfill.">
<opted-out are suppressed — do not re-reach.>

---
Privacy: works from classifications + brand-safe intent summaries only; no
contact info and no raw reply text in this memo (convention 7 + connection-flow
privacy invariants). Drafts are suggestions you send — the matchmaker relays;
contact is exposed only after a creator's own yes, never by this skill.
Provenance: Creatorland Data MCP · list_connections + get_connection_status
(read-only, free) · <date>.
Credits used this run: 0 (all connection reads are free; no creators reached).

Honesty rules

• Read-only — never reaches. This skill calls list_connections and get_connection_status only. It does NOT call request_creator_connection, so it never spends the 10-credit reach charge and never messages a creator.
• Drafts are suggestions you send, not sent messages. The skill writes the reply; the brand approves and sends it, and the matchmaker relays it. Say so — never imply the draft went out.
• Never promise a yes. "Interested" means the sequence got a positive reply and the matchmaker is brokering an intro — it is not a confirmed booking. Frame outcomes as the creator's to make.
• Privacy (convention 7 + connection-flow invariants). No contact info and no raw reply text in the deliverable, regardless of source. Work from the classification + brand-safe intent summary; if a summary looks like it carries contact details, drop to the bare classification. The connection's existence + status is brand-safe; a creator's address is not.
• No invented comp. Drafts fill comp/budget only from a brand-supplied or corpus rate band; otherwise leave [brand to confirm]. Never quote a figure the brand didn't authorize.
• Don't re-reach the declined. Opted-out creators are suppressed; suggesting re-contact is a harassment + sender-reputation risk. Acknowledge and move on; backfill comes from a fresh cast, not a re-poke.
• Graceful, never error. No replies, no entitlement, or an empty list each resolve to a plain-language "nothing to triage / not on your plan" — never a stack trace.

Credit footprint

thorough: 0 credits — every call here (list_connections, get_connection_status) is free; the skill reaches no one · thrifty: identical (0 credits) — thrifty only trims the status filters to the buckets the user asked for (e.g. questions only), which makes no cost difference since the reads are already free.