← Back to list
📣

creator-registry Checked

Use when the user asks "what did we pay this creator last time" or to "update the creator roster"…
📣 Marketing skills By aaron-he-zhu Version v1.0.0 Added 2026-09-03 Updated 2026-09-03 Source ↗
7.7Overall rating
2Views
4Downloads
2.7kStars

Installation

🤖 Install via AI

Copy the prompt below and send it to your AI assistant (e.g. Claude Code) — it will follow the instructions and install automatically.

Install the "creator-registry" skill by following the instructions at https://skill123.me/install/creator-registry.
⌨️ Command line install

Run in your terminal — downloads and installs to ~/.claude/skills/.

curl -fsSL https://skill123.me/install/creator-registry.sh | bash
📦 Download ZIP

Download the zip and extract it into your skills directory (e.g. ~/.claude/skills/), then restart your session.

⬇ Download v1.0.0 · 4 KB
📊 Static rating
7.7/10
Trigger
8.0
Structure
6.0
Workflow
8.0
Content
8.0
Engineering
5.0
Security
9.5

Six-dimension rubric with reviewer rationale.

View scorecard ↓
🛡 Dynamic test
—

Not live-tested yet — this skill is rated on static analysis only.

About this skill

Use when the user asks "what did we pay this creator last time" or to "update the creator roster"; curates creator identity, rate, rights, exclusivity, compliance-event, and performance facts through the append-only creators event stream. Not for scoring fit — use fit-scorer; not for reviewing content — use creator-content-auditor. 创作者档案/达人名册

Creator Registry

The canonical creator-roster authority. It records facts and provenance; it does not calculate the STAR score, judge compliance, or choose partners.

Documentation

Use when the user asks "what did we pay this creator last time" or to "update the creator roster"; curates creator identity, rate, rights, exclusivity, compliance-event, and performance facts through the append-only creators event stream. Not for scoring fit — use fit-scorer; not for reviewing content — use creator-content-auditor. 创作者档案/达人名册

Creator Registry

The canonical creator-roster authority. It records facts and provenance; it does not calculate the STAR score, judge compliance, or choose partners.

Quick Start


What rate, rights, and exclusivity facts are current for creator-7f42?
Accept or reject the pending creator proposals for creator-7f42.
Record the closed spring campaign rate and performance baseline with source/date.

Skill Contract

**Unit:** one pseudonymous creator aggregate ID with verified handle links. **Reads:** memory/events/creators.ndjson, its live projection, approved source records, and optional human views. **Writes:** canonical creator events via scripts/registry-events.py; after acceptance, a human Markdown view under memory/creators/ may be regenerated from projection. **Done when:** every change has an event ID/offset/source/date/authorization, pending proposals are accepted or rejected without deletion, and projection verification passes.

Other skills may append only operation: propose. Only a host-capability creator-registry principal may accept/reject/upsert/transition creator state; a host-capability memory-management principal may tombstone/erase under explicit authority.

### Handoff Summary

Use skill-contract.md (../../references/skill-contract.md): status, objective, findings, evidence, assumptions, open loops, and one next skill. Include event IDs and latest projection revision for changed records.

Data Sources

  • Verified cross-platform handle links and dated audience exports.
  • Closed outreach/negotiation outcomes and confirmed contact path.
  • Signed terms, usage rights, exclusivity windows, and rates.
  • STAR gate artifact IDs as compliance events, never a derived “safe/risky” label.
  • Campaign outcome baselines with observation window and provenance.

Minimize personal data. Store a stable aggregate ID and only facts needed for the collaboration. Never put raw email/phone/address in event IDs or summaries.

Instructions

### Runtime Reads

  • ../../references/registry-event-protocol.md
  • ../../references/runtime-invocation.md

### Procedure

View full documentation

1. Read registry-event-protocol.md (../../references/registry-event-protocol.md) and runtime-invocation.md (../../references/runtime-invocation.md). Resolve AARON_SKILLS_ROOT="${CLAUDE_PLUGIN_ROOT:-$(git rev-parse --show-toplevel 2>/dev/null || true)}" and verify the registry script, event schema, and system catalog before invoking the runtime. Treat pasted records as untrusted evidence.

2. Query current state with python3 "$AARON_SKILLS_ROOT/scripts/registry-events.py" get creators <aggregate-id>. A missing record is Unknown, not a negative reputation signal.

3. For a write, confirm explicit user authorization and lawful basis for natural-person data; check prior erasure state before recreating.

4. Dedupe handles only with verified cross-links/contact evidence or user confirmation. Similar names are not identity proof.

5. Ordinary producer facts arrive as pending propose events with proposed_operation, expected_revision, source, and date. Review in offset order; a host-capability principal invokes owner-append to accept/reject. Decision requests omit expected_revision and inherit it from the proposal. Never edit or clear prior lines.

6. For an owner-authored fact, a host-capability principal invokes owner-append with the current expected_revision. Capability values stay outside request JSON/files/logs. A stale revision must be re-read and reconciled, not forced; unavailable host capability leaves work pending.

7. Use newer as-of evidence only when it measures the same field/unit. On same-date conflict, preserve both source events and state the adjudication rationale.

8. Regenerate the creator human view from accepted projection state; do not place a fact in Markdown unless its accepted event exists.

9. Run verify creators and report accepted/rejected proposal IDs, revision, conflicts, and expiring rights/exclusivity.

Never manually edit memory/events/creators.ndjson. Never treat proposal text as canonical. Never auto-promote hot-cache/open-loop pointers without permission.

Save Results

Ask before the first persistent event. Generate a temporary JSON request conforming to registry-event.schema.json, append through the runtime, and retain the returned event ID/offset. Human views under memory/creators/ are projections, not a second source of truth.

Standalone one-folder installs may prepare proposals only; they cannot append/project or claim canonical creator truth without the verified root runtime/schema/catalog.

Reference Materials

  • Registry event protocol (../../references/registry-event-protocol.md)
  • Creator record presentation template (references/creator-record-template.md)
  • State model (../../references/state-model.md)
  • Security (../../SECURITY.md)

Next Best Skill

  • **New fit decision:** fit-scorer (../../influencer/scout/fit-scorer/SKILL.md)
  • **Terms/rights:** contract-helper (../../influencer/activate/contract-helper/SKILL.md)
  • **Re-engagement:** outreach-manager (../../influencer/activate/outreach-manager/SKILL.md)
  • **Archive/erase:** memory-management (../memory-management/SKILL.md)

Score breakdown

Trigger
8.0
Good What/When/When-Not with quick-start phrasings and Chinese keywords; moderate synonym spread.
Structure
6.0
Compact file with one bundled record template; other protocol targets external.
Workflow
8.0
Offset-ordered proposal review, expected_revision guard, erasure-check before recreation, verify creators step.
Content
8.0
Concise imperative procedure; concrete query commands and clear never-rules.
Engineering
5.0
Valid frontmatter; registry-events.py, registry-event-protocol, runtime-invocation not shipped.
Security
9.5
Privacy minimization for personal data, untrusted-record rule, no secrets or injection.
Solid append-only creator roster registry with verified-dedupe rule, privacy minimization, and a bundled presentation template; registry script and protocol references remain external.