| name | goblin-recon |
| description | Goblin Recon router skill. Routes user commands to the smallest useful workflow: Social Pulse, Clip Mine, Clip Vault, Competitor Scout, Brand Gate, Email Hook, or Carousel Generator. |
Goblin Recon — Router Skill
Goblin Recon finds what is trending, finds the source, and finds the moment.
Use this operating shape for every request:
Router -> Workflow -> Tools -> Normalized Data -> Score/Gate -> Human Gate -> Memory
Do not run every source or workflow by default. Pick one primary workflow, run the fewest reliable tools, then stop with a clear next action.
Router
| User intent | Workflow | Output |
|---|
| Trends, hooks, formats, ideas | Social Pulse | Ranked content opportunities |
| Source videos or timestamps | Clip Mine | Editor-ready clip briefs |
| Retrieve/update prior clips | Clip Vault | Clip lists, briefs, status updates |
| Competitor analysis | Competitor Scout | Competitor report |
| Brand/voice fit | Brand Gate | Pass / revise / shelve |
| Email hooks or drafts | Email Hook | Scored subject/opening variants |
| Carousel or social image | Carousel Generator | Rendered local assets |
If the request mixes workflows, run the smallest sequence and state it. Example: run full scan = Social Pulse first, then Clip Mine only for the 2-3 strongest candidates.
run full autonomous scan means use approved public/local sources and local tools without repeated confirmations. It still cannot reveal secrets, bypass access controls, publish/send, create paid services, delete important files, deploy, or commit without explicit approval.
Output Direction
Before brand-facing output, ask once:
- Audience: B2C, B2B, or Both?
- Destination: Faceless Instagram, personal brand, client work, internal use, email/outbound, or other?
- Tone: professional, casual, edgy, warm, wry, reflective, analytical/data-driven, bold, or platform-native?
If skipped, default to Both / Faceless Instagram / professional and state that default.
Skip this only for Clip Vault retrieval, direct brand checks on user copy, or manual scans where direction is already supplied.
Workflows
Social Pulse
Purpose: ideas, trends, hooks, formats, carousels, blogs.
Flow:
Scan mode -> source collection -> social_intake -> scoring -> Security Rail -> report
Scan modes:
- Fast Scan: YouTube, Reddit, Tech News, Product Hunt, approved/public X. Avoid fragile IG/TikTok unless requested.
- Deep Social Scan: Instagram/TikTok public surfaces first, then X/Reddit/Tech News validation.
- Signal Scan: X/public, Hacker News, GitHub Trending, ArXiv, then Reddit if public. Last 6 hours.
- Manual Assisted Scan: user supplies URL/screenshot/caption/handle; normalize and score.
Every Social Pulse report starts with ## Decision, includes category, sources, publication dates, effort, confidence, vault check, fallback angle, AI search potential, hooks, formats, creators, and next step. Save to vault/reports/YYYY-MM-DD-social-pulse.md.
Clip Mine
Purpose: source videos and 15-60 second moments.
Flow:
Trend/topic -> Source Hunter -> transcript -> Clip Vault dedup -> Moment Finder -> Brand Gate -> Security Rail -> Human Gate
Rules:
- YouTube first, then public Instagram/TikTok, then podcast platforms.
- Clip duration: 15-60 seconds, natural sentence start/end.
- Validate timestamps with
goblin_recon.tools.clip_extractor.
- Check duplicates with
goblin_recon.tools.clip_store before recommending.
- Save only approved/shelved metadata and short excerpts, not full raw transcripts.
- Briefs use
templates/clip-mine-brief.md and include Instagram, LinkedIn, YouTube Shorts variants.
Clip Vault
Purpose: local memory for approved, shelved, and production-status clips.
Use:
scripts/query_clips.py list --status approved
scripts/query_clips.py list --query "topic"
scripts/query_clips.py brief <clip_id>
Storage: vault/clips.db, vault/briefs/, memory/trend-history.md.
Competitor Scout
Purpose: public competitor intelligence. Keep separate from Social Pulse and Clip Mine.
Use templates/competitor-report.md. Every claim needs a public source URL and date. Run Security Rail before delivery.
Brand Gate
Load memory/brand-rules.md and config/brand-voice.yaml before outward content. Run goblin_recon.tools.brand_gate when feasible.
Default recommendation when uncertain: shelve.
Email Hook
Use skills/email-hook/SKILL.md and goblin_recon.tools.email_gate for outbound subject lines, openers, and short drafts. Ask Output Direction first.
Carousel Generator
Use skills/carousel-generator/SKILL.md and goblin_recon.tools.carousel_renderer.
Flow:
brief/topic -> brand memory -> platform spec -> slide manifest -> render -> QA -> vault/carousels
Render locally. Replicate is optional; no token means gradient fallback.
Source Rules
Social extraction ladder:
Approved API/public feed -> public browser extraction -> manual assisted input
Never bypass login, paywall, captcha, robots.txt, rate limits, or platform restrictions. Never use personal employee accounts. If blocked, mark access_status: blocked and move on.
Normalize social records through goblin_recon.tools.social_intake before scoring.
Required fields: platform, creator, url, published_date, views, likes, comments, caption, hook, format_type, topic, category, why_it_is_trending, can_genx_adapt_this, confidence, access_status.
Canonical Tools
| Tool | Use |
|---|
api_search.py | Exa/Tavily search |
api_extract.py | Firecrawl/ScrapeGraph extraction |
youtube_tool.py | transcripts and public video metadata |
clip_extractor.py | clip URL/timestamp validation |
scoring.py | velocity, lifecycle, source diversity |
social_intake.py | normalize social observations |
clip_store.py | save/find/update local clips |
brand_gate.py | brand voice checks |
email_gate.py | email quality scoring |
carousel_renderer.py | carousel/social image output |
MCP source of truth: mcp.json.
Final Gate
Before user-facing reports, clip briefs, competitor claims, publish instructions, or brand copy, run skills/security-rail/SKILL.md mentally or explicitly.
Security Rail decision: APPROVE, REVISE, SHELVE, or NEEDS HUMAN REVIEW.
Deliver only safe/revised output. If shelved, do not recommend it.
Output Rules
- Start reports and briefs with
## Decision.
- No fabricated sources. No URL = do not include.
- All data needs publication date when applicable.
- Category required: Latest AI News, Controversial/Polarizing, Upgrade/Democratization, or Analytical/Deep-dive.
- Recommendation before evidence.
- Phone-scannable output.
- Always suggest the next step.
