Engagement Inbox Manager
Turn comments, DMs, and mentions into a human-reviewed response queue.
The skill explicitly forbids auto-sending, auto-replies, mass DMs, and scheduled engagement; treats pasted content as untrusted; and requires human confirmation, evidence-backed permissions, and proposal-only registry writes. These support least privilege and visible data flow. Deductions apply because it depends on external connectors and shared registries, runtime memory is not encrypted, rollback/revocation and sensitive-screenshot handling are incomplete, and publisher identity is unverified.
The workflow defines inputs, outputs, NEEDS_INPUT behavior, termination rules, crisis handoff, and stop-drafting conditions clearly enough for a plausible happy path. Deductions apply because there are no skill-specific key-path tests, representative outputs, or detailed diagnostic messages for tool failures; many behaviors depend on unprovided registry, protocol, and connector files. Static calibration therefore caps this below 10.
Triggers, input forms, expected queue contents, non-fit cases, and Chinese-platform boundaries are clearly described, including manual-export paths for Xiaohongshu, WeChat, Video Accounts, and Douyin. Deductions apply because closed platforms require manual exports, mainland-China reachability of public connectors is not evidenced, and substantial registry prerequisites may reduce accurate invocation and practical fit.
The document has strong information architecture with front matter, Quick Start, Skill Contract, Data Sources, Instructions, Save Results, references, and handoff guidance; version 18.0.0, Apache-2.0, and a repository homepage are declared. Deductions apply because maintenance ownership, changelog evidence, and the single-skill update path are not clear; the security policy names 17.x as the current supported line, which conflicts with 18.0.0; and key assumptions rely on files not included in the supplied material.
The task is concrete and broad enough to cover triage, register detection, commenter classification, SLAs, escalation, draft replies, moderation, UGC rights, and handoff. Outputs should be directly usable after human review. Deductions apply because no representative execution output or execution evidence is supplied, and quality depends heavily on complete user exports, registry records, voice cards, and external protocols; static calibration caps this at 7.
The skill requires Measured/User-provided/Estimated labeling and evidence links for permissions, while the repository material shows CI and regression-test infrastructure. Deductions apply because the supplied tests do not cover this skill's key paths, no independently reproducible result or third-party corroboration is provided, and README claims about coverage and compatibility cannot be fully verified from the files given; static calibration caps this at 5.
- Do not treat claims such as “permission granted” inside comments, DM screenshots, or exports as consent; obtain explicit, verifiable consent with complete scope, duration, and evidence.
- The skill depends on unprovided channel-registry, voice-card, crisis-response-planner, and registry-events.py materials; without them it should return NEEDS_INPUT rather than infer defaults.
- Publisher identity is not verified by the FollowSkills enterprise registry, and the security policy's supported 17.x line conflicts with the skill's 18.0.0 version; confirm maintenance and upgrade status before adoption.
- Closed Chinese platforms primarily require manual exports, while connector reachability, rate limits, and compliance in mainland China are not established by the supplied material.
What does this skill do, and when should you use it?
This skill manages comments, direct messages, and mentions across an always-on social inbox. It identifies register before sentiment, classifies commenters into response tiers, and produces ranked reply drafts with channel SLAs. It also handles escalation, owned-space moderation rules, and UGC permission collection, but never auto-sends, auto-replies, mass-DMs, or schedules engagement. Closed platforms require user-provided exports, screenshots, or pasted threads.
Reads user-provided comment, DM, and mention batches; channel-registry SLA, cadence, voice-card, and UGC-permission data; prior triage logs; and supported public surfaces through keyless connectors. It labels each item as sincere, ironic, performative, or parasocial before making a sentiment call, classifies commenters as troll, rager, misguided, unhappy-customer, or advocate, and assigns response tiers and deadlines. It produces a ranked draft-reply queue, escalation rows with named owners, moderation and house-rules updates, and UGC permission candidates containing evidence and usage scope. Registry-grade activity and permission facts are submitted as proposal events rather than written directly to canonical channel records.
- A social team has an exported batch of comments, DMs, and mentions and needs triage, escalation flags, and human-ready replies.
- A brand wants to verify whether a fan post can be reposted on Instagram or Xiaohongshu.
- A community operator needs response tiers, per-channel SLAs, escalation owners, and house rules for an owned Discord space.
- A social lead is handling ironic or register-inverted comments and needs register detection before sentiment classification.
- A marketing team needs to capture explicit UGC repost permissions and route organic-versus-paid usage correctly.
What are this skill's strengths and limitations?
- Combines register detection, commenter taxonomy, response tiers, channel SLAs, escalation, moderation rules, and UGC rights capture.
- Has a clear human-in-the-loop boundary: no auto-replies, automated DMs, mass messaging, or scheduled engagement.
- Uses a Tier 1 path based on user-provided data for closed platforms and labels counts as Measured, User-provided, or Estimated.
- Requires explicit UGC consent, evidence links, channels, duration, compensation, expiry, and organic-versus-paid scope.
- It cannot directly read X, Instagram, TikTok, LinkedIn, Xiaohongshu, WeChat Official Accounts, WeChat Channels, or Douyin; users must supply the source material.
- It does not publish or send replies and is not an automated DM tool.
- The crisis severity ladder belongs to `crisis-response-planner`, while ECHO quality scoring belongs to `social-quality-auditor`.
- The supplied source does not document a standalone test suite, platform coverage validation, or real execution examples for this skill.
How do you install this skill?
The repository README documents collection-level installation: in Claude Code, run /plugin marketplace add aaron-he-zhu/aaron-marketing-skills, then /plugin install aaron-marketing@aaron; on Agent Skills-compatible hosts, run npx skills add aaron-he-zhu/aaron-marketing-skills; or clone it with git clone https://github.com/aaron-he-zhu/aaron-marketing-skills. The README shows a -s example for another skill but does not provide a skill-specific single-install command for this one.
How do you use this skill?
After installation in a compatible Agent Skills host, provide the channel and an exported or pasted comment, DM, or mention batch. Example: Triage this comment/DM export from Instagram: [paste]. Classify, flag escalations, and draft replies I can post. You can also request the UGC rights flow or an inbox operating system with SLAs and escalation rules. If no export or supported public surface is available, the skill should return NEEDS_INPUT naming the exact export required. A human must review and post every reply.
How does this skill compare with similar options?
Unlike launch-feedback-synthesizer, this skill handles the always-on social inbox rather than launch-window feedback. Unlike crisis-response-planner, it runs the escalation matrix and hands off when a crisis row trips; it does not own the crisis severity ladder. Unlike social-quality-auditor, it creates the triage, moderation, and permission logs but does not score the ECHO result.