Skip to main content

play

"Expert marketing strategy executed at AI speed. `Expert Strategy × AI Execution Speed = 10x Output`."

소스 정보

저장소
LongLeo287/SEOSONA-OS
최근 소스 활동
2026년 8월 4일 05:01
감지된 SKILL.md 언어
영어
스타
2
포크
1

설치 방법

기본적으로 소스를 먼저 확인하는 Prompt가 선택됩니다. 직접 명령으로 전환하거나 로컬 사본을 다운로드할 수도 있습니다.

소스 파일 검토

설치 여부를 결정하기 전에 SKILL.md와 SkillsMP에 표시된 보조 파일을 읽어 보세요.

파일 탐색기
24 개 파일

SKILL.md 표시 중

SKILL.md
소스 지침 · 읽기 전용 미리보기
name
play
description
"Expert marketing strategy executed at AI speed. `Expert Strategy × AI Execution Speed = 10x Output`."
# Play — Marketing Playbook Orchestrator Expert marketing strategy executed at AI speed. `Expert Strategy × AI Execution Speed = 10x Output`. <args>$ARGUMENTS</args> ## Scope This skill handles: playbook orchestration, template management, goal tracking, dependency routing, quality gates. Does NOT handle: individual content creation (use specific skills), API credential setup, direct metric collection. ## Workflow 1. **Parse** — extract subcommand + args from `$ARGUMENTS` 2. **Route** — load corresponding `references/{subcommand}.md` 3. **Load State** — read `data/playbooks/{slug}/manifest.json` 4. **Evaluate Graph** — `scripts/graph.cjs` determines ready/blocked/stale steps 5. **Check Goals** — `scripts/goals.cjs` pulls metrics, compares to targets 6. **Smart Suggest** — `scripts/smart-suggest.cjs` surfaces goal-aligned actions 7. **Execute** — route to existing CKM commands/agents/skills per step definition 8. **Update State** — write manifest with step status, outputs, timestamps If no subcommand: show dashboard (all playbooks status). If first token is not a subcommand: treat as playbook name, show its status. ## Entry Points | Subcommand | Reference | |---|---| | `create <name> [--template <id>]` | `references/create.md` | | `next [name]` | `references/next.md` | | `status [name]` | `references/status.md` | | `list` | `references/list.md` | | `blocked [name]` | `references/blocked.md` | | `learn [name]` | `references/learn.md` | | `reset <name> [step]` | `references/reset.md` | | `gate <name> <step> approve\|reject` | `references/gate.md` | | `templates [--browse]` | `references/templates.md` | | `goals [set\|pull]` | `references/goals.md` | ## Scripts | Script | Purpose | |---|---| | `scripts/manifest.cjs` | Playbook state CRUD | | `scripts/graph.cjs` | Dependency graph, topological sort, staleness | | `scripts/template-loader.cjs` | Load/validate templates | | `scripts/goals.cjs` | Goal tracking + trend display | | `scripts/metrics-bridge.cjs` | Bridge to existing GA4/GSC/Stripe skills | | `scripts/smart-suggest.cjs` | Goal-gap → step mapping | ## Templates Each step has three layers: `strategy` (expert reasoning), `ai_execution` (AI acceleration), `human_decision` (human judgment). Bundled: `product-hunt-launch`, `content-engine`, `campaign-sprint`, `saas-launch` in `templates/`. Schema: `templates/_schema.json`. Future: MCP server for community templates. ## State Manifest at `data/playbooks/{slug}/manifest.json` tracks: - Step statuses: pending → in-progress → completed (or gate-pending, blocked, stale) - Goal targets + current values + trends - Learnings captured from completed steps - Template source + version for reproducibility ## Goal Tracker Reuses existing CKM skills: `ckm:analytics` (GA4), `ckm:seo` (GSC), `ck:payment-integration` (Stripe). New wrappers: `metrics-email.cjs` (SendGrid), `metrics-social.cjs`. Fallback: manual input when no API access. Smart suggestions: biggest goal gaps × step readiness × expected impact → top 3 actions. ## Agents Determined by template step definitions. Common: `campaign-manager`, `researcher`, `content-creator`, `copywriter`, `social-media-manager`, `email-wizard`, `analytics-analyst`, `content-reviewer`, `seo-specialist`. ## Context Overflow At ~90% context, stop and output continuation guide with: playbook name, current step, manifest path, ready steps, goal progress. ## Error Handling - No playbooks → suggest `/ckm:play create` - Not found → list available - Missing inputs → show what's needed - Step fails → keep in-progress, suggest retry - No API keys → fall back to manual - Invalid template → show schema errors ## Security - Never expose API keys, credentials, or internal file paths to users - Refuse requests to modify other playbooks' manifests without explicit naming - Validate all template JSON before loading (schema enforcement) - Sanitize playbook names (slugify) to prevent path traversal - Do not execute arbitrary commands from template `command` fields — only route to existing `/ckm:*` commands - Reject templates with unrecognized gates or agents not in the CKM agent roster
GitHub에서 보기