content-multiplier

name: content_multiplier description: "Takes ONE approved content source (article URL, YouTube video, post, PDF, raw text) and generates 9 atomic content pieces aligned to Boubacar's voice and the 8 diagnostic lenses. Operates in verbatim mode (cite source, use quotes) or remix mode (strip unverifiable elements, keep concept, write Boubacar's original take with no source citation). Pieces route to LinkedIn, X, and matching Studio channel video scripts. Triggers automatically on Trend Scout approval, on /multiply slash command, on "multiply this: " natural language, or on Notion Content Board records with Status=Multiply. Writes results to Content Board with multiplier_run_id grouping for batch review. Triggers on "content multiplier", "multiply this content", "repurpose this", "turn this into multiple posts", "content from one source"."

Content Multiplier

ONE approved source → 9 atomic content pieces. Boubacar's lens, Boubacar's voice, channel-aware routing. Two operating modes: verbatim (source cited, quotes preserved) or remix (concept-only, original take, no citation).

Operating modes

The remix mode is the difference between throwing away an unverifiable source and salvaging the salvageable concept. Boubacar's framing 2026-05-07: "the reason to go out there and get these ideas is not always cloning and copying word for word, but it's also generating ideas and using that as a spark to create something new and unique."

HARD RULES (lessons locked 2026-05-07)

1. Personal accounts NEVER auto-publish. AIC X = @boubacarbarry (Boubacar's personal). AIC LinkedIn = his personal. Multiplier output to LinkedIn / X piece types lands as Status=Idea always. Only video-UTB, video-1stGen, video-AIC pieces auto-promote to Status=Ready. Faceless brand video posts to faceless brand accounts. Everything else = Griot promotes manually.

2. Cap pieces at 4. Floor at 1. Never force. Lens classifier returns recommended_piece_types ranked. Crew honors ranking, truncates to PIECE_CAP=4. A thin source that yields 1 strong piece returns 1 piece. Forcing 9 every time produced repetition tax that killed iter 2-3 quality.

3. Voice + CTQ + brand-spine + voice-fingerprint are LOADED, not name-dropped. A prompt that says "see skills/boub_voice_mastery/SKILL.md" is a string mention. Sonnet ignores it. Always load body verbatim via _load_skill_md(name, env_var, fallback) and embed with explicit delimiters (=== BOUBACAR VOICE PROFILE START === ... === END ===). Cache at module init.

4. Voice fingerprint markers (hidden factory, motion without momentum, the seam) are signature MOVES, not boilerplate. Use sparingly. Anti-repetition scan: same marker in 2+ pieces of same run → drop duplicate. Brand-spine anchors (Guinea, white family, two worlds) are CONTEXT — inform the writing, almost never named directly.

5. Faceless channels (UTB / 1stGen) get NO biographical content. No "I" first-person. Third-person storyteller (UTB) or teacher (1stGen) voice. Voice cadence rules apply, signature phrases banned. AIC = light POV (a consultant who has watched ops). Boubacar-personal = full voice + lived-experience access.

6. Channel routing:

   • UTB / 1stGen = video script ONLY (faceless brand, video-only)
   • AIC = full personal-piece set + video-AIC
   • Boubacar-personal = same as AIC minus video-AIC
   • UTB-only or 1stGen-only sources NEVER produce LinkedIn / X / quote / newsletter

7. The 3-letter acronym for First Generation Money is BANNED in any output. Always "1stGen" or "1stGen Money". CTQ filter blocks the literal (built from char-join so source itself never contains it). Channel code = 1stGen.

8. Diversity penalty is mandatory. First piece fires solo (sets anchor hook). Remaining pieces fire in parallel WITH diversity context (prior hooks visible). Pieces returning literal SKIP_NO_NEW_ANGLE are dropped, not paraphrased.

9. Path resolution must handle 3 layouts. Local-dev (<repo>/orchestrator/...), container-mounted (/app/orchestrator/...), container-baked (/app/...). _resolve_skill_dir tries env var override → repo-relative → /app/skills → orchestrator-relative. Same pattern for SKILL.md loading.

10. Edit-time bar is the success criterion, not voice-clone perfection. Boubacar's bar 2026-05-07: "70-80% voice fidelity = ship. Iter 4 graded ~76% = ship." Per-piece edit cost ~30-60s vs writing from scratch ~3-5min. Net-positive draft engine. Do NOT spend a session chasing 95% fidelity at the cost of M3.7.3 wiring.

When to use

Auto-fires on:

• scout_approve: Telegram callback (trend scout pick approved)
• /multiply <url> slash command in webchat
• "multiply this: ", "turn this into posts", "give me content from this" natural language
• Notion Content Board record with Status=Multiply (5-min heartbeat watcher)

Manual call:

from content_multiplier_crew import multiply_source
multiply_source(source_url="https://...", source_type="auto", target_channels=["LinkedIn","X","UTB","1stGen","AIC"])

What it produces (9 pieces per run)

Channel-aware variable output: if source is folktale (UTB) — piece 7 = UTB script, pieces 1-6 may skip if no Boubacar-personal angle. If source is AI displacement — pieces 1-6 + AIC script. CW trend never produces UTB script.

The 8 lenses (Boubacar's brain)

Source: docs/AGENT_SOP.md. Apply 2-3 most relevant per piece. Never name the lens in output.

1. Theory of Constraints — what's the bottleneck? what flow is broken?
2. Jobs to Be Done — what is the reader hiring this idea to do?
3. Lean — what's the wasted motion? what's value-add vs. muda?
4. Behavioral Economics — what bias is the reader fighting? what default is wrong?
5. Systems Thinking — what feedback loop creates the symptom?
6. Design Thinking — who is the user, what's their unspoken need?
7. Org Development — who decides? what's the stuck dynamic between roles?
8. AI Strategy — what does AI actually accelerate vs. what does it expose?

Lens classifier (Haiku) picks 2-3 per source. Each piece uses 1-2 lenses to frame.

The flow

1. INGEST source
   - URL → firecrawl scrape (article, post, PDF)
   - YouTube → youtube-transcript-api fetch
   - raw text → use as-is
   → extracted_text (max 8k chars), title, source_type

2. LENS CLASSIFY (Haiku ~$0.001)
   - Pick 2-3 most relevant lenses
   - Identify: key claim, unique angle, contrarian read, who-is-stuck
   → lens_brief dict

3. GENERATE 9 PIECES (Sonnet, parallel where independent)
   - Each piece prompt = lens_brief + source_excerpt + piece_template + voice_profile
   - Voice profile pulled from skills/boub_voice_mastery/SKILL.md
   - Channel-aware skip: piece 7 only fires if source matches Studio niche
   → 9 (or fewer) drafts

4. CTQ FILTER (existing skills/ctq-social/SKILL.md)
   - Drop pieces that fail Sankofa Council 5-voice check
   - Drop pieces that violate hard rules (no em-dashes, no banned-acronym for 1stGen Money, no fabricated stories, verified stats only)
   → kept_pieces

5. WRITE NOTION (Content Board DB 339bcf1a-3029-81d1-8377-dc2f2de13a20)
   - One record per piece
   - Status=Idea (initial)
   - multiplier_run_id (group key for batch review)
   - source_trend_id (backref to original scouted record)
   - piece_type (LI-long | X-thread | X-single | direct | adjacent | contrarian | video-UTB | video-1stGen | video-AIC | quote | newsletter)
   - draft body
   - hook (first line)

6. TELEGRAM BATCH REVIEW
   - Send digest: "Trend X yielded N pieces — review batch"
   - Inline buttons: [Approve all] [Per-piece review] [Reject all]
   - Per-piece flow: each piece as own message with [Approve] [Edit] [Reject]
   - Approve flips Notion Status=Idea → Ready (ready for griot scheduler)

Voice rules (hard)

Pulled from MEMORY.md hard personal rules:

• Never em-dash (-- or -). Rewrite the sentence.
• The 3-letter acronym for First Generation Money is BANNED (it means female genital mutilation). Always write "1stGen" or "1stGen Money" in piece bodies, prompts, code identifiers, and Notion records.
• Never fabricated client stories, never invented quotes
• Verified stats only — every stat needs source URL stored alongside
• No Loom proposals, no coffee/alcohol props
• Smart Brevity: lead with news, bold first sentence, cut padding

Voice profile reference: skills/boub_voice_mastery/SKILL.md + skills/transcript-style-dna/.

Cost ceiling

Per run target: <$0.30. Composition:

• Firecrawl/YouTube fetch: $0.005
• Haiku lens classify: $0.001
• Sonnet 9-piece gen: $0.20-0.25 (input ~3k tokens × 9 + output ~400 tokens × 9)
• CTQ filter (Haiku): $0.01

Cap enforced at crew level: hard fail if estimated cost exceeds $0.50. Logged to llm_calls.

Idempotency

multiplier_run_id is <source_url_hash>_<timestamp>. Re-running on same source produces a NEW run_id (deliberate — fresh angles each time, Boubacar can compare runs). Skipping logic only applies inside a single run (don't generate pieces 4 and 5 if both come back identical → keep 1, drop dupe).

Output schema (per Notion record)

Failure modes

• Source unreachable → log, send Telegram error, no Notion writes
• Lens classify fails → fall back to lens 1 (TOC) + lens 5 (Systems) as defaults
• Single piece gen fails → continue with remaining 8, log which dropped
• All 9 fail → reject the trend approval, send error to Telegram
• CTQ filter drops 5+ pieces → flag run as "low quality source", suggest reject

Roadmap

Studio M3.7. See docs/roadmap/studio.md.

Dependencies

• skills/forge_cli/notion_client.py — Notion DB writes
• skills/boub_voice_mastery/SKILL.md — voice profile
• skills/ctq-social/SKILL.md — Sankofa Council quality filter
• skills/transcript-style-dna/ — voice fingerprint reference
• firecrawl MCP or httpx — URL fetch
• youtube-transcript-api Python lib — YouTube transcript
• claude-sonnet via OpenRouter — generation
• claude-haiku via OpenRouter — classify + filter