Skip to content

106 social media skills for AI agents - strategy, writing, video, design, platform growth, publishing, and analytics. Works with Claude, Cursor, OpenClaw, Hermes & 40+ agents.

3.5 KB
Raw Blame

Descript — integration & connection guide

What it is: the text-based talk-content editor (podcasts, interviews, talking-head, courses) in this stack. Descript edits; WoopSocial publishes the social exports; podcast RSS hosting is separate. Skill: skills/descript/. Plans/credits changed hard in Sept 2025 — verify-quarterly at descript.com/pricing and the in-app meters; treat pre-2025 reviews as stale.

Connection layers

  1. Public API (2026, open beta) + MCP: programmatic access to projects, media imports, and Underlord actions — including via Model Context Protocol connections — so an agent can trigger filler/silence removal, clip flagging, show notes, and imports where connected. Actions consume AI credits; batch prompts to conserve.
  2. In-app (human executes): no connection? The agent writes the cut list (transcript-anchored), Underlord prompts, the Overdub decision, and the export/repurpose plan; the human executes on desktop (macOS/Windows) or web (near-parity in 2026; cloud sync; cloud-dependent — no offline editing).
  3. Publish handoff: exports (episode, clips, captions) → scheduling-and-queue → WoopSocial (upload media → attach → validate → create post; per-platform required fields validate atomically — see woopsocial.md). RSS/podcast hosts are outside WoopSocial — the human publishes there.

Plan gates (≈, sources conflict — verify in-app)

  • The Sept 2025 overhaul: transcription-minutes → media minutes (ALL imported media counts) + AI credits metering Underlord, Studio Sound (~10/use), Overdub, Eye Contact, Green Screen, avatars. Top-ups expire in 12 months; documented bill-shock pattern and no mid-cycle proration (G2, attributed).
  • Tiers (figures conflict): Free ≈ 60 media min/1hr + watermark · Hobbyist ≈ $16–24 · Creator ≈ $24–35 · Business ≈ $50–65 (annual ~25–35% cheaper). Never quote a tier as fact. SOC 2 Type II; project info confidential per Descript.
  • Credit-aware workflow: import only what you'll edit · Studio Sound once per source · batch Underlord · Overdub words/phrases only · watch the meter.

The voice-consent spine (the reason this file exists)

Overdub is consent-verified and own-voice-only by design — identity verification blocks cloning anyone else's voice. That is the stack's model for voice cloning generally: never clone a guest, competitor, or public figure; never type words into a real person's mouth; paid tiers include commercial use of your own consented clone (verify terms); AI-disclosure for synthetic speech where platform/region requires (EU AI Act; C2PA). Craft limit: words/phrases sound like you; paragraphs drift synthetic — re-record.

Known limits + routing

Transcription ~92–95% clean English (accuracy pass mandatory; ~75–85% with noise/accents/jargon; ~23 languages) · not a beat-synced short-form editor (→ CapCut) · not a cinematic finisher (→ Resolve/Premiere) · not a transcription-at-scale/compliance service · Automatic Multicam needs separate speaker tracks (record that way on purpose) · Eye Contact patches an off-camera read, not delivery (→ talking-head-and-piece-to-camera).

Hard lines

Human verifies by ear + approves every cut (and owns the final cut of anyone else's words) · interview edits preserve meaning + clip context · own-voice-only cloning, disclosed where required · accuracy pass before captions ship · never state tiers/credits as fact without in-app verification · WoopSocial does not edit media or host podcasts.