1. Routing matrix
| Incoming | Skill / action |
|---|---|
| YouTube link | skills/transcript/SKILL.md → transcript |
| X/Twitter link | skills/twitter/SKILL.md → tweet content |
| Voice/audio (.ogg) | skills/groq-voice/SKILL.md → transcription |
| Web link | skills/markdown-new/SKILL.md (fallback: skills/chromium/SKILL.md) |
| Market request | skills/market-data/SKILL.md → data + indicators |
| Screenshot/image | Vision models (describe content, extract text) |
Extract content via routing matrix before command routing.
2. Command routing
Explicit command → execute:
| Command | Action |
|---|---|
| "write a post" / "make a post" | Post pipeline (§3) |
| "record idea" | Save to ideas storage |
| "this is my stance" / "my position" | Update stance/core |
| "save as reference" | Save to library (is_reference: true) |
| "save" | Save to sources |
No explicit command → ASK:
"Got it, this is [brief summary]. What should I do with it? -- write a post / record as idea / save as material / update position / just for reference"
Never guess intent.
3. Post pipeline (10 steps)
- Channel. Crypto/macro → main channel. Ambiguous → ask.
- Idea. From user or from ideas storage -- suggest 2-3. After approval → mark status in_progress.
- Persona Lock. Load from channel profile. Override per-task only if explicitly requested.
- Context (required sources):
- TOV (Tone of Voice): load from your TOV storage
- Channel profile: load from channel profiles storage
- Library (3-5 posts): load from library storage
- Stance (crypto/macro): load from stance/core storage
- Local mirrors as fallback if storage unavailable
- Pre-draft brief. 1 thesis, 2-4 supporting points, sources. Show to user or go straight to draft (simple tasks).
- Draft. Opus subagent. Pass: persona lock + TOV + library + task. Save to drafts storage.
- Fact-check (BEFORE humanizer!). Verify: numbers not from API, historical dates, quotes, names. Links only from sources, never invented.
- Humanizer. 3 passes per
skills/content-engine/references/humanizer-patterns.md. DO NOT touch: facts, numbers, levels, position, categoricality. - Show. Draft + fact-check to user. Silently, without listing steps. Apply edits verbatim, don't soften. Ask: "One-time edit or add to TOV?"
- Publish. ONLY on command. After → offer to save to library. Update idea → status: published.
4. Fallbacks
| Situation | Action |
|---|---|
| Library empty | Write by TOV + stance, warn: "Library empty, style may be imprecise" |
| Storage unavailable | Local mirrors, mark [unverified] |
| Stance conflict | Ask user: update stance or one-time exception? |
| Cross-post | One draft → adapt for each channel (TOV + profile for each) |
5. Market report
On "what's the market doing" / "market overview":
- Use
skills/market-data/SKILL.mdto collect data (BTC, RSI, F&G, OI, funding, liquidations, SPY, Fed). - Apply stance from stance/core storage.
- Format: price → indicators → macro → conclusion through user's position.
6. YouTube
Scripts, descriptions, timecodes -- details in skills/content-engine/references/youtube-guide.md.
YouTube scripts go through steps 3-9 of the pipeline (persona lock, context, brief, draft, fact-check, humanizer, show).
Clarify format with user: Shorts / 5-7 min / 15-20 / 30+. Outline → approval → full script.
7. Two idea banks
| Bank | Purpose |
|---|---|
| Content ideas | Posts, videos |
| Project ideas (status=pipeline) | NOT for content |
8. TOV update (every 5 posts)
- Get learnings from storage -- edits for the period.
- Get library -- recent originals.
- Patterns (minimum 2 repetitions) → suggest 2-5 rules → after "yes" → update TOV storage.
9. Rights and rules
- ideas, sources, drafts -- without asking.
- library -- only after "add to library".
- TOV -- only on explicit request.
- Draft ALWAYS show to user.
- Publish -- ONLY on command.
- Library = main style source.
- Humanizer pass -- MANDATORY between fact-check and show.
- Persona Lock -- MANDATORY before draft.
- Storage commands: full reference in
skills/content-engine/references/storage-commands.md.