| name | influencer-pack |
| description | Pre-configured creator-economy vertical pack for the ai-brain-starter substrate. Ships typed-memory categories for audience, DMs, content, collabs, and creator revenue; connector configs for Instagram, TikTok, YouTube, Substack, X, LinkedIn, Stan Store, Stripe, Patreon; retention defaults aligned with Meta + TikTok + YouTube data-handling expectations; the idea engine, an evidence-grounded content-idea generator that learns the creator's taste from every discard; decision-audit patterns for sponsored-disclosure and evidence-grounding. Use when onboarding a creator, personal-brand operator, or content-engine team that needs the substrate to come pre-shaped to the work rather than starting from a blank vault. |
| trigger | /influencer-pack |
| argument-hint | init | status | rebuild [--platform <name>] |
/influencer-pack
A pre-configured pack that turns the empty substrate into a creator-economy ready system in one install. The pack ships typed-memory categories that match how audience, content, DMs, and collabs actually move; connectors for the platforms creators already use; retention defaults that map to Meta and TikTok and YouTube data-handling expectations; and decision-audit patterns for sponsored-content disclosure and brand-deal terms.
Why this exists
A blank install of the substrate forces every creator to invent the same vocabulary on day one: what is an audience segment, what is a DM intent, what is a collab stage, how do I attribute revenue across sponsorship + subscription + product + course + speaking + affiliate. The vocabulary is not novel. Creators run roughly the same primitives, against roughly the same set of platforms.
This pack ships the primitives so the creator can spend day one on what is actually theirs (their voice, their audience-relationship history, their content engine) instead of re-deriving the category structure.
What this pack sets up
Run /influencer-pack init and the pack writes:
| Layer | What ships | Where it lands |
|---|
| Schema | 12 typed-memory categories with frontmatter contracts | schema/typed-memory-categories.md |
| Connectors | 16 platform connector specs (Instagram, TikTok, YouTube, Substack, X, LinkedIn, Stan Store, Stripe, Patreon, more) | connectors/*.md |
| Idea engine | The content-generation spine: bucket structure, discard loop, compounding taste profile, pre-filter interface | idea-engine/*.md |
| Retention | Per-category retention defaults aligned with platform terms of service and creator-economy norms | retention/defaults.md |
| Decision audit | Sponsored-content disclosure and evidence-grounding patterns | decision-audit/*.md |
Nothing is auto-applied to a live install. The pack stages drafts under drafts/ and prints the path; the operator reviews and accepts before merging into the production memory layer.
The idea engine
The idea engine is the pack's spine. It is what turns the typed-memory layer from an archive into a working content engine, and it is the piece the rest of the pack orbits.
It does two things a generic AI idea generator does not:
- Every idea is grounded in real audience evidence. No idea is proposed without a verbatim quote from a real comment, DM, or piece of the creator's own content behind it. An idea that cannot cite real evidence is dropped, not softened. The enforcing pattern is
decision-audit/evidence-grounding.md.
- It learns the creator's taste. Every idea the creator discards is logged with a reason and compounded into a
taste-profile, a weighted, confidence-scored model, not a rolling reject list. Later runs read the profile and stop proposing what the creator keeps rejecting. The engine improves with use, and the profile is the creator's switching cost.
The engine reads content-piece, dm-conversation, and audience-question; it writes content-idea, idea-discard, and taste-profile. The full mechanism (bucket structure, generation flow, the discard loop, the compounding-profile math, the prompt-cache layout) is in idea-engine/mechanism.md. The noise filter that keeps adoration and bot text out of the generation context is in idea-engine/pre-filter.md.
/content-engine is the creator-facing skill that runs the engine.
Required companion skills
The influencer-pack assumes these substrate skills are also installed in ~/.claude/skills/. The connectors call into them directly. Install them alongside the pack:
| Companion skill | Purpose | Source |
|---|
ingest-youtube | Pulls YouTube transcripts into the vault as queryable markdown so podcast appearances, video essays, and competitor analysis flow into the same typed-memory layer the rest of the pack queries. Required for the /repurposing-engine and /voice-fingerprint-update skills. | Bundled in this repo at skills/ingest-youtube/ |
ingest-slack | Pulls Slack channels (creator community, agency channel) into the vault for cross-platform context. | Bundled in this repo at skills/ingest-slack/ |
humanizer | Voice-fingerprint pass on AI-generated content drafts before publish so the creator's voice does not flatten across the engine output. | https://github.com/adelaidasofia/humanizer |
graphify | Knowledge graph extraction across the typed-memory layer so the creator can see content + DMs + collabs as one navigable surface. | Bundled in this repo at skills/graphify/ |
Skill bundle (creator-facing commands)
| Skill | What it does | Trigger |
|---|
/dm-closer | Drafts DM responses in the creator's tone, qualifies fan vs prospect vs brand-collab, escalates high-value contacts to manual review. | Always-on |
/content-engine | Runs the idea engine: proposes evidence-grounded content ideas across three audience buckets and learns from every discard via the taste profile. Also turns approved ideas and anchor recordings into platform-native drafts. See idea-engine/mechanism.md. | Weekly batch |
/weekly-creator-report | Monday-morning rollup: revenue by source, top content by reach plus revenue, audience growth, collab pipeline status. | Cron at 7am Monday |
/collab-pipeline | Tracks inbound brand asks plus outbound creator-to-creator collabs with stage, terms, next-step. | Always-on |
/voice-fingerprint-update | Re-trains the written-voice fingerprint on the last 30 days of approved content. Reads from External Inputs/YouTube/ (via ingest-youtube), the Substack archive, and the IG caption history. | Monthly cron |
/repurposing-engine | Turns a long-form recording (podcast, YouTube video, keynote) into platform-native cuts: short-form video, newsletter post, X thread, LinkedIn post. Reads transcripts from the typed-memory layer, which means ingest-youtube must run first on the source URL. | Per-recording |
/launch-automation | Course or cohort or paid-product launch sequence with email plus DM plus Story plus Reel cadence. | Per-launch |
Companion visual surface (cover art, thumbnails, post visuals)
Creator content lives or dies on the visual layer. The pack assumes one of two image-gen MCPs is available; the engine routes to whichever matches the visual register:
| Use case | Primary | Backup / alternate |
|---|
| Hero / OG / blog cover for a content idea (photorealistic register) | nano-banana (Gemini 3 Pro Image) — structured deliverables with alt-text | midjourney-mcp for the MJ aesthetic |
| Cinematic / painterly / fantastical cover (Midjourney register) | midjourney-mcp imagine → wait_for_task → upscale U1-U4 on the keeper | nano-banana (cheaper, different aesthetic) |
| Reverse-engineer a competitor's hero aesthetic | midjourney-mcp describe (image-to-prompt) — pass image_url, get 4 candidate prompts | manual prompt-writing |
| Composite reference moodboard into one new cover | midjourney-mcp blend (2-5 image URLs in, 4-up grid out, optional dimension square|portrait|landscape) | manual moodboard |
| Iterate on a chosen direction | midjourney-mcp variation V1-V4 (regenerates a new 4-up grid off a chosen position) | imagine with tweaked prompt |
| Sharp-edges discipline before any image-gen call | Anti-keyword-soup, anti-text-in-prompt, positive framing, physics-aware relationships, character continuity, Perfect Prompt formula, intent → framing mapping | inline in each MCP's README |
Cost discipline. midjourney-mcp (PiAPI-backed; github.com/adelaidasofia/midjourney-mcp) gates every cost-incurring call on a daily USD cap (MIDJOURNEY_MCP_DAILY_USD_CAP, default $5/day, calendar-day reset at the configured UTC offset). Per-call estimates: imagine $0.04, variation $0.04, upscale $0.01, describe $0.005, blend $0.04 (fast mode). For creator-scale batches (10+ covers/week), pre-flight by raising the cap or switching to PiAPI Host-Your-Account ($8/seat/mo flat, set cap=0). nano-banana (Gemini 3 Pro Image) is free of per-call cost — default for high-volume work; reach for midjourney-mcp when the AESTHETIC specifically calls for it.
Init flow
- The operator runs
/influencer-pack init.
- The pack reads the existing substrate state and reports any conflicts (categories already defined, connectors already configured).
- Drafts land at
drafts/<timestamp>/ under the pack root with the proposed schema, connector configs, and retention defaults.
- The operator reviews the drafts. They can edit in place or run
/influencer-pack init --regenerate <file> to redraft a single file.
- When the operator accepts, the pack writes to the production memory layer and registers the companion skills as installed dependencies.
Status check
/influencer-pack status reports:
- Which typed-memory categories are live versus draft
- Which connectors are configured versus pending
- Retention rules applied vs default
- Companion skill install state (
ingest-youtube, humanizer, graphify)
- Most recent ingest run per connector
- Pending items in the creator's review queue (DMs, collab proposals, sponsored-content disclosures)
Rebuild
/influencer-pack rebuild --platform <name> re-pulls the connector spec from the latest version in this repo, diffs against the local install, and stages an upgrade plan. Use after a major platform API change (Meta Graph version bumps, YouTube Data API revisions).
Companion-skill detail: ingest-youtube
The /repurposing-engine and /voice-fingerprint-update skills both depend on YouTube transcripts being in the typed-memory layer. The flow:
- Operator runs
/ingest-youtube <url> (single video) or /ingest-youtube <channel> --days N (recent uploads).
ingest-youtube calls yt-dlp, cleans the VTT subtitle file, and writes External Inputs/YouTube/<channel-slug>/<YYYY-MM-DD>-<video-slug>.md with full transcript and metadata.- The next time
/voice-fingerprint-update runs, it picks up the new transcripts and re-trains.
- The next time
/repurposing-engine runs against that source URL, it reads the cached transcript instead of re-fetching.
This is why the pack lists ingest-youtube as a hard dependency, not an optional integration. Without it, the creator-facing skills cannot reach the source content.
See connectors/youtube.md for the platform-specific notes (manual subtitles vs auto-captions, channel-mode pagination, language preference defaults).
