kelly-campaigns

name

kelly-campaigns

description

Outbound email-marketing desk (App-in-Skill) for building segments, drafting campaigns, newsletters, and sequences, running pre-send deliverability and subject-line QA, and approving every send before it is scheduled. Structured around the SEND discipline — Setup, Engage, Nurture, Deliver — with an email-quality-auditor gate (EQS score + SHIP/FIX/BLOCK verdict). Use when the user invokes $kelly-campaigns or /kelly-campaigns, or mentions email marketing, campaigns, newsletters, drip/welcome/win-back sequences, broadcasts, segments, subject-line A/B tests, deliverability (SPF/DKIM/DMARC/spam score/inbox placement), or wants to review and approve marketing email before it is sent. This is OUTBOUND marketing to a subscriber list, distinct from kelly-email inbox triage (incoming mail). 出站邮件营销：策划分群、起草营销活动 / 新闻邮件 / 邮件序列，发送前做可送达性与主题行质检，人工审批后再排期发送。

Kelly Campaigns

Overview

Use this skill as Kelly's outbound email-marketing operator. It keeps a file-backed App-in-Skill dashboard over an email program: audience segments, drafted sends (campaign broadcasts, newsletter issues, and sequence steps), pre-send deliverability and subject-line QA, and post-send performance. The skill builds segments, drafts email copy, runs deliverability and quality checks, and prepares sends; the human reviews, edits, and approves each send in the app before anything is scheduled or sent.

This is outbound marketing to a subscriber list. It is distinct from kelly-email, which triages an incoming inbox. Keep them separate: kelly-email is about mail you received; kelly-campaigns is about mail you send to many people.

Default interaction mode: App UI. Unless the user explicitly asks for chat-only handling, check onboarding/config, refresh or regenerate the local campaign snapshot, start/reuse the local app with app/start.sh, and give the actual local URL. Use chat-only mode only when the user says "纯聊天", "chat only", "不要打开 UI", or similar; in that mode present numbered sends (Send #1) and take verdicts in the conversation.

This skill is an implementation of the App-in-Skill pattern — a Codex/agent skill paired with a small local companion UI for review and approval. See the spec paper: https://mr-kelly.github.io/research/app-in-skill-specification-for-pairing-agent-skills-with-a-local-companion-ui.pdf.

The SEND Discipline

Work is organized around four phases. Every send is tagged with the phase it belongs to, and the app groups by phase.

• Setup — the sending foundation: authentication and deliverability, audience segments, list growth, list hygiene.
• Engage — the creative: email copy, subject lines, HTML render / dark-mode, dynamic personalization.
• Nurture — the lifecycle: sequences (welcome, drip, win-back), newsletter cadence and monetization, preference/frequency management, reactivation.
• Deliver — getting it into the inbox: send experiments (A/B), inbox-placement monitoring, cold outbound, and the pre-send quality gate.

Capability taxonomy (16 sub-skills)

Phase	Sub-skills
Setup	deliverability-qa · list-segment-builder · list-growth-designer · list-hygiene-monitor
Engage	email-creative-builder · subject-line-lab · email-render-builder · dynamic-content-personalizer
Nurture	email-sequence-designer · newsletter-monetization-planner · preference-frequency-manager · reactivation-specialist
Deliver	send-experiment-designer · inbox-placement-monitor · cold-outbound-sequencer · email-quality-auditor ⛩ (the quality gate)

The quality gate — email-quality-auditor ⛩

Before any send can be scheduled, it passes the SEND framework, which produces an EQS (Email Quality Score, 0–100) and a SHIP / FIX / BLOCK verdict:

• S — Sender & auth: SPF, DKIM, DMARC pass; correct from-identity.
• E — Engagement risk: segment quality, expected open/complaint behavior, re-permission for cold/lapsed lists.
• N — Not spammy: spam score, trigger words, link/image balance, working unsubscribe + physical address.
• D — Deliverability: inbox readiness against policy floors, IP/domain warm-up, render and dark-mode.

Verdicts: SHIP (ready to schedule), FIX (deliverable but revise first), BLOCK (hard stop — e.g. failing DKIM or a spam score above policy). The gate never sends; a human still approves. A block verdict or deliverability.risk === "high" refuses scheduling even if a stale approve decision exists.

App UI Screenshots

Overview Send calendar plus list health — subscribers, bounce, churn, and complaint rates.	Campaigns Draft and approval queue across campaigns, newsletters, and sequence steps.
Deliverability Pre-send QA — SPF/DKIM/DMARC, spam score, and the EQS SHIP/FIX/BLOCK gate.	Performance Open, click, and unsubscribe rates by campaign.

Boundary

• The skill may build segments, draft sends, run deliverability/quality checks, validate schemas, and write local handoff files.
• The app reads and writes local files only. It must never send email, call an ESP, mutate a list, or perform any external side effect.
• Sending is always approval-required (outbound + volume). Real scheduling/sending is performed by the configured ESP by the skill, only after the user approves the specific send in the app or in chat. scripts/execute_decisions.ts only records handoff operations in execution_report.json; it performs no sending itself.
• Treat subscriber data as sensitive. Do not commit config.local.json, env files, app/.data/, exports, or subscriber lists.

First Run And Onboarding

On invocation, check app/.data/onboarding.json and private config readiness. If onboarding is absent/incomplete, guide setup before doing real marketing work.

Private config priority:

1. KELLY_CAMPAIGNS_CONFIG=/absolute/path/to/config.json
2. skills/kelly-campaigns/config.local.json
3. ~/.config/kelly-campaigns/config.json
4. skills/kelly-campaigns/config.example.json as template only

Env priority:

1. Existing environment variables
2. KELLY_CAMPAIGNS_ENV_FILE=/absolute/path/to/.env
3. Repository root .env
4. skills/kelly-campaigns/.env.local
5. ~/.config/kelly-campaigns/.env

Ask for non-secret setup details only: operator profile (name, role, company, timezone), brand (name, homepage, unsubscribe URL), ESP provider name, from-identities (from-name/from-email/reply-to and when to use each), segments, sending policy (approval-required, daily/hourly caps, min inbox readiness, max spam score), risk keywords, and style/tone. Never ask the user to paste secret values into chat — the ESP API key belongs only in a local env file, referenced from config by *_env name.

When setup is complete and the user confirms, write app/.data/onboarding.json:

{
  "completed": true,
  "completed_at": "ISO timestamp",
  "config_version": "1"
}

Local App

Start the dashboard with:

skills/kelly-campaigns/app/start.sh

The app uses local HTTP on 127.0.0.1, preferring port 3210 (override KELLY_CAMPAIGNS_UI_PORT), then falling through 3210–3999 if occupied. The launcher reuses a running instance only when /api/state proves it is the same app (app: "kelly-campaigns").

Required app views (hash routes):

• #/overview: send desk. Human-attention counts, upcoming sends for the next weeks, list health (subscriber count, bounce/complaint/churn rates, avg open/click), and a SEND-phase breakdown.
• #/campaigns and #/campaigns/<send_id>: the review queue over drafted sends in workflow states needs_review, changes_requested, approved, done, blocked. Each item shows a stable row ref (Send #1), type + phase + quality-gate verdict badges, the segment and audience size, deliverability risk, subject + preview text, an editable body draft, an A/B subject picker when variants exist, a Review note textarea, and Approve / Request changes / Block buttons that write to decisions.json. Sidebar workflow filters and SEND-phase chips both apply. The queue is read-only while agent.lock exists.
• #/deliverability: pre-send QA table — SPF/DKIM/DMARC pass flags, spam score, inbox readiness, and the SEND verdict per send, so weak auth or spammy copy is caught before scheduling.
• #/performance: open/click/unsub/bounce by sent campaign.
• #/settings: sanitized config summary — operator, brand, ESP + secret readiness, from-identities, segments, sending policy, data provider name, and onboarding state. Never expose secret values.

Keep the sidebar workflow filters (All / Needs Review / Approved / Done / Blocked) as the primary nav, plus the views above.

Demo mode:

• ?demo=1 opens a deterministic mock program for documentation and screenshots.
• ?demo=overview, ?demo=campaigns, ?demo=deliverability, ?demo=performance, and ?demo=detail select named mock scenes; detail deep-links to a send detail.
• lang=en or lang=zh forces UI chrome language for screenshots.
• Demo API responses must never read or write files under app/.data/ or any private config.

UI language: support English and Chinese chrome with Auto default. Keep subject lines, body copy, segment names, and drafts in their original language.

File Contract

Read references/campaigns-schema.md before editing the app, scripts, or any generated campaign JSON.

Primary local files:

• app/.data/campaigns_snapshot.json: normalized program snapshot (segments, sends, list_health, metrics, warnings) generated by the skill/scripts.
• app/.data/decisions.json: user verdicts, review notes, edited bodies, and chosen A/B variants keyed by send id, written by the app.
• app/.data/agent_tasks.json: queued agent work — sends in changes_requested with the user's comment. The skill polls this to pick up revisions.
• app/.data/execution_report.json: latest ESP handoff results written by scripts/execute_decisions.ts.
• app/.data/onboarding.json: onboarding completion marker.
• app/.data/agent.lock: temporary lock while the skill is generating or executing. The app rejects decision writes while it exists.
• config.local.json: private operator configuration, ignored by git.

Use scripts/validate_ui_schema.ts before relying on a snapshot in the UI. The app may show an empty setup state when no snapshot exists.

Normal Workflow

1. Detect mode. Default to App UI.
2. Load private config through the config helpers. If only config.example.json exists, enter onboarding.
3. When Kelly asks for a campaign, newsletter, or sequence: acquire app/.data/agent.lock, build/refresh the relevant segment, draft the send(s) into sends[] with status: "needs_review", the correct type and SEND phase, a clear reason, risk badges, deliverability check, subject variants when A/B is intended, and run the email-quality-auditor gate to attach quality_gate (EQS + verdict). Recompute metrics and list_health, validate with scripts/validate_ui_schema.ts, then release the lock.
4. Start/reuse the UI and report the URL so Kelly can review the queue, deliverability, and quality gates.
5. Poll app/.data/agent_tasks.json for changes_requested items. Re-draft each one according to the user's comment, re-run the gate, set it back to needs_review, and clear the task.
6. On "schedule approved sends": re-read decisions.json, re-check the lock, and run scripts/execute_decisions.ts --apply to record schedule_send / ab_test operations in execution_report.json. Then perform the actual scheduling/sending through the configured ESP with the approved, possibly user-edited body and chosen variant, one send at a time, and mark each done afterward.
7. Never schedule a send without an explicit approve decision, never schedule one whose gate is BLOCK or whose deliverability risk is high, and never re-schedule sends already recorded as scheduled/sent.

Safety Defaults

• Treat every outbound send as approval-required (outbound + volume). Money offers, compliance-sensitive copy, cold outbound, and high send volume raise the bar further.
• A BLOCK verdict or high deliverability risk is a hard stop; fix authentication or copy before the send is eligible.
• Store only the minimum content needed for review; keep raw subscriber lists and PII out of the snapshot (segments carry names + sizes, not rows).
• Redact the ESP API key and any credential-like strings from logs, reports, and UI state; expose only boolean readiness for configured env vars.
• Keep stable ids (send_id, segment_id) and ref numbers so repeated updates and executions are idempotent.
• If decisions and the snapshot disagree (missing send, stale ref), stop and regenerate rather than guessing.