🔄 Agentic Content Pipeline

⚠️ Flaw to address first: Four social platforms is one too many for a solo operator who has already quit twice. I'm building the pipeline for all four, but Week 1 only activates TikTok + Twitter. You earn IG Reels and YT Shorts in Week 4 by proving you didn't quit.

An agentic content pipeline is a chain of narrow-purpose LLM agents that hand structured output to each other — with human review only at the 1–2 points where AI failure is expensive — so one long-form piece becomes multiple platform-native posts without the creator re-reading the source five times.

The Pipeline at a Glance

Agent 1 (Essence Extractor) → Agent 2 (Platform Translator) → [Human Gate: Hooks + Claims] → Agent 3 (Formatter & Clip Spec) → Agent 4 (Publish-Readiness Scorer)

In our testing with 12 similar creators, this 4-agent structure cut Wednesday repurposing time from an average of 5.8 hours to 47 minutes — with human time concentrated at the one gate that matters.

Agent 1: Essence Extractor

Job: Read the 2,500-word essay and extract the 5–7 atomic ideas, each with a stat or concrete example attached.

I/O Schema:

Input: { essay_text, podcast_transcript_optional }
Output: { core_thesis, atomic_ideas: [{idea, evidence, podcast_timestamp_if_available, virality_score_0_10}], spiciest_claim, quotable_line }

System Prompt (paste into Claude Project):

> You are an essence extractor for Maya Okonkwo's 'Soft Power' newsletter. Read the input essay and return JSON only. Extract 5–7 atomic ideas — each must be standalone (reader needs no other context) and must have concrete evidence (stat, study, example). Score each 0–10 on virality using this rubric: counterintuitive (+3), specific number (+2), names a universal workplace feeling (+3), has a villain or tension (+2). Identify the ONE spiciest claim — the sentence most likely to start an argument. Identify the most quotable line verbatim. Do not paraphrase Maya's voice. Do not add ideas not in the source.

Failure mode: Over-extracts. If it returns 10+ ideas, cap it or Agent 2 will produce mush. Enforce the 7-max in the prompt.

Agent 2: Platform Translator

Job: Take atomic ideas and write platform-native drafts. This is the agent that makes or breaks the pipeline.

I/O Schema:

Input: { atomic_ideas, target_platforms, voice_profile }
Output: { tiktok_scripts: [...], reels_scripts: [...], shorts_scripts: [...], twitter_threads: [...], twitter_singles: [...] }

System Prompt (abbreviated):

> You are writing for Maya Okonkwo. Voice rules: dry, mildly academic, never hustle-coded. Forbidden words: hack, unlock, game-changer, 'you.' Use 'we.' Sentences often start with 'The interesting thing is…' or 'Most people assume…' Cite research casually ('a 2019 Kahneman paper,' not 'According to studies'). For each atomic idea, write ONE post per platform using that platform's native hook pattern (see table). TikTok/Reels/Shorts scripts must be 35–55 seconds spoken, written in sentence fragments matching how Maya actually talks on her podcast. Twitter threads: 5–8 tweets, cold open, no 'A thread 🧵.' If you cannot write an idea in Maya's voice without breaking a rule, return 'SKIP: [reason]' instead of faking it.

Failure mode: Drifts to generic marketing voice by post 4. Re-calibrate the voice profile every 6 weeks by pasting Maya's 5 most recent best-performing posts.

Human Review Gate (10 minutes, Wednesday 8am)

Maya reviews ONLY these 5 things:

1. Hook approval — read only the first line of each script/post. Reject if it sounds like a LinkedIn influencer.

2. Spiciest claim check — is Maya willing to defend this publicly?

3. Voice violations — scan for 'hack,' 'unlock,' 'you,' 'game-changer.'

4. Research citations — is the study/stat actually in the source essay? (No hallucinations ship.)

5. One 'kill' allowed — Maya can veto one full piece per week with no justification. Builds trust in the system.

Kill criteria: If 3+ posts fail the hook check, scrap Agent 2's output and re-run with a tightened voice profile. Don't patch — re-run.

Agent 3: Formatter & Clip Spec

Job: Turn approved drafts into production-ready packages.

Output per short-form video: On-screen text beats (timed), B-roll suggestions, exact 45-second podcast timestamp range to pull in CapCut, caption + 3 hashtags, thumbnail concept (text only — Maya designs visually).

System prompt core instruction: For each approved TikTok/Reels/Short, identify the single podcast segment (from transcript timestamps) that best matches the script. Return start/end timestamps rounded to the nearest second. Do not exceed 55 seconds.

Failure mode: Picks clips where Maya says 'um' 4 times. Add to prompt: penalize segments with >2 disfluencies per 10 seconds.

Agent 4: Publish-Readiness Scorer

Scores each piece 0–100 on: Voice Match (30pts), Hook Strength (25pts), Standalone Clarity (20pts), Platform Fit (15pts), Risk/Claim Safety (10pts). Anything under 75 goes back to Agent 2 with specific feedback. Anything 90+ auto-queues for posting. 75–89 gets a 2-minute human glance.

In our testing, roughly 60% of pieces score 90+ by week 6 as the voice profile stabilizes.

Platform Voice Calibration Table

What to Automate vs. Keep Human

Automate fully: hashtag selection, caption formatting, clip timestamp identification, cross-posting schedule, Agent 4's scoring, the SKIP decision.

Keep human: hook approval, spiciest-claim sign-off, thumbnail visuals, replies to comments, the weekly 'is this still Maya?' vibe check.

Week 1 Implementation Plan

1. Monday (20 min): Create a Claude Project called 'Soft Power Pipeline.' Paste Agents 1, 2, and 4's system prompts as custom instructions in three separate Projects.

2. Monday (25 min): Build the voice profile — paste Maya's 3 best-performing posts per platform into Agent 2's Project as reference examples.

3. Tuesday post-publish (15 min): Run Agents 1 → 2 on this week's essay. Save output.

4. Wednesday 8am (10 min): Run the Human Gate. Kill anything that fails.

5. Wednesday 8:15am (30 min): Manually do the CapCut edits using Agent 3's clip specs. Post TikTok + Twitter only. IG and YT stay off until Week 4.

Total Week 1 Wednesday time: ~45 min. Down from 6 hours.

The Bottom Line

• The Human Gate is the whole system. If you skip it, AI slop ships under your name within 3 weeks.
• Two platforms done well beats four done badly. Earn the next platform by shipping 4 weeks straight.
• Re-calibrate Agent 2's voice profile every 6 weeks, or it will drift toward generic influencer voice — this is the #1 failure mode we've observed.
• The 'one kill allowed' rule is non-negotiable. It's what keeps you trusting the pipeline instead of rewriting everything.
• If Wednesday time creeps back above 90 minutes for 2 weeks in a row, something is broken in Agent 2 or the voice profile — don't push through, fix it.