Social AI Manager: Swipe Intake first, then the Autopilot Studio

1. Problems & risks first

Blunt list. Red items block autonomy; amber items are unverified assumptions the design depends on.

• Compliance approval blocks full autopilot. Under NAIC Model 570 every Reel is an "advertisement," and the producer who made it is personally responsible. Recruiting, testimonials, anything naming a carrier or product, annuities and Section 125/Longtable cannot auto-publish. They need Chris's one-tap approval, and carrier-named content also needs a logged carrier pre-approval ID. No AI can grant that.
• Recruiting content is where enforcement already happened. The FTC sent IMO Family First Life a cease-and-desist over agents' income posts and said the IMO is responsible for its participants' claims. No income figures, ever. Typical-earnings data is unanswered.
• TikTok requires GHL (or another audited aggregator). TikTok will not audit a home-built "upload to my own account" tool, so unaudited apps can only post SELF_ONLY to private accounts. Whether Chris's GHL plan exposes Social Planner API scopes is unverified. Fallback: Zernio.
• config/facts.md is empty ("(none yet)"). Model 570 requires every statistic to have a named, recent source. Until Chris approves facts, no script can use a single number.
• The marketing bible interview sheet is unanswered. 222-Agent-Marketing-Bible v0.5 has 36 questions, including "pick your top 3 hooks." It stalled Sept 27. Without it the script agent writes generic content, and generic finfluencer content drifts toward phrases regulators target.
• Local Whisper conflicts with the "no fans" rule. Chris asked to "transcribe locally with Whisper." Running Whisper on his Mac is exactly the CPU/fan load he wants to avoid, and the Cowork sandbox also runs on his Mac. Cloud Whisper costs pennies a month, so it is the default here. Local whisper.cpp is an opt-in overnight toggle. Details in Swipe Intake.
• Unverified platform facts: Instagram API publish cap (Meta's page says both 100 and 50 per 24h; design assumes 50 and reads content_publishing_limit); Business Discovery 2026 field list (competitor view counts may not come back); Trial Reels via trial_params untested on his account; YouTube "unverified projects are private-locked" claim (2026 docs say no, test once); Submagic B-roll and silence-cut API parameters; the 140 to 160 WPM pace (uncited).
• Swipe Intake practical limits. Instagram doesn't let you save most other people's Reels to Photos, so screen recording is the realistic capture. Screen recordings include IG's UI overlay, so the breakdown must ignore UI chrome. Worker request bodies cap around 100 MB, so files larger than that need a two-step presigned R2 upload. iCloud may keep files as cloud-only placeholders, so the watcher must force a download first.
• Copyright. A private transcript kept for analysis is low risk. Republishing it, or a close paraphrase, is not. The system borrows structure and hook type only and rejects any script with a run of 6+ consecutive words matching the reference.
• No AI voice or avatars. Meta, TikTok and YouTube label realistic synthetic voices and faces. An "AI info" tag would make the post "about AI," which breaks Chris's own rule. The Oct 19 "AI voiceover videos" idea conflicts with this.
• Vendor prices are mostly from competitor comparison pages. No vendor virality score has independent validation. Claude 2026 API prices were not retrieved, so they are estimates.

2. What already exists

The file sweep found no roadmap or build for a social media manager. One working piece, and a lot of stalled inputs.

Swipe Intake (primary research lane)

Chris sends Reels he likes. The cloud does the rest. His Mac uploads at most, and only on route (b).

1. Capture. Save or screen-record the Reel on iPhone (Control Center → Screen Recording) or Mac (Cmd+Shift+5). Optional: note the source handle and view count shown on screen.
2. Send, by one of three routes:
   • (a) iOS share-sheet Shortcut "Send to TFG Swipe" (recommended). Share → Send to TFG Swipe → optional note ("@handle, liked the cold open, 1.2M views"). The Shortcut POSTs file + note to https://swipe.<worker>/swipe with header Authorization: Bearer <SWIPE_TOKEN>. The Worker writes to R2 tfg-swipe and enqueues a job. The Mac is not involved. Over ~100 MB: the Shortcut first asks the Worker for a presigned R2 URL, then PUTs the file there.
   • (b) iCloud Drive folder "TFG Swipe Inbox". A tiny launchd agent (same pattern as linkedin-daily) uses WatchPaths on ~/Library/Mobile Documents/com~apple~CloudDocs/TFG Swipe Inbox. It wakes only when a file lands, runs brctl download if the file is a placeholder, curl-uploads it to the Worker, and moves it to Sent/. Upload only, near-zero CPU, no transcription on the Mac.
   • (c) Claude via Dispatch or chat. Attach the video (and a note). Claude posts it to the same Worker endpoint, so the pipeline is identical.
3. Transcribe. Default: cloud Whisper. Groq whisper-large-v3-turbo or OpenAI gpt-4o-mini-transcribe, about $0.003/min or less, so roughly 60 one-minute refs a month cost under $0.25. Word timestamps are kept for pacing math.
4. Breakdown. Gemini watches the actual video (not just the transcript), so it sees on-screen text, the visual hook and the cuts. It is told to ignore Instagram UI overlay. Output JSON:
   • Hook in the first 3 seconds: verbal, visual and on-screen text hooks, plus hook type (question, contrarian, sourced stat, mistake/warning, story open, POV/skit, list, authority, visual interrupt, direct callout)
   • Structure/beats with timestamps
   • Pacing: words per minute (from ASR timestamps), cuts per 10 seconds
   • Caption and on-screen text style
   • CTA
   • Format: talking head, green screen, skit, list or story
   • Length
   • "Why it works"
   • Transferability to TFG, scored 1 to 5
5. Log. Write to Supabase swipe_refs (file, route, note, source, private transcript) and swipe_breakdowns (the analysis), tagged by theme, hook_type, format and niche. If Chris supplies view counts (and optionally the account's typical views), store an outlier note: views ÷ typical, flagged at 3× or more.
6. Generate. Claude writes 3 original TFG scripts per reference in Chris's voice:
   • one consumer life-insurance script
   • one recruiting script (yellow lane)
   • one flex script (Section 125/Longtable, or a TFG value angle)
   Rules: the writer receives the breakdown, not the raw transcript. It borrows structure and hook type only, never wording. An n-gram check then compares each draft to the reference transcript (lowercase, punctuation stripped) and rejects any draft with 6 or more consecutive matching words, regenerating up to twice. The report's pgvector similarity block (0.85 cosine) runs as a second check. Every surviving script then goes through the Compliance Gate.
7. Deliver. Scripts that pass land in the teleprompter folder/queue (S042-H1.txt, S042-H2.txt, S042-BODY.txt in the shared Drive folder) with lane flags. Chris gets a Discord notification in the same channel pattern as linkedin-daily, with the breakdown summary, transferability score and Approve buttons for yellow-lane scripts.

Illustration: one reference, three original scripts (structure borrowed, wording not)

Reference pattern (abstracted): mistake/warning hook, green screen over a headline, 3-point list, comment-keyword CTA, 32 s.

Samples follow house rules: no numbers needing a fact source, no guarantees or income claims, no em dashes, soft CTA. Recruiting sample still routes to the yellow lane.

3. Framework diagram

Swipe Intake is stage 0 and primary. The scrape lane is dotted because it is optional and later. One human job: recording.

flowchart TD
  subgraph S0["0. SWIPE INTAKE (primary)"]
    A1["Chris saves or screen-records a Reel"] --> A2{"Send route"}
    A2 -->|"a. Share Sheet Shortcut (recommended)"| W["Cloudflare Worker /swipe + secret token"]
    A2 -->|"b. iCloud folder + launchd watcher"| W
    A2 -->|"c. Claude via Dispatch or chat"| W
    W --> R2[("R2 tfg-swipe")]
    R2 --> TR["Transcribe: cloud Whisper default, local whisper.cpp opt-in"]
    TR --> GB["Gemini breakdown: hook, beats, pacing, CTA, format, why it works, transfer 1-5"]
    GB --> LG[("Supabase swipe_refs + swipe_breakdowns")]
  end
  subgraph S0B["Optional scrape lane (later)"]
    SC["Sandcastles, Apify, IG Business Discovery, YouTube Data API"]
  end
  LG --> RS["1. Research: patterns + weekly outlier brief"]
  SC -.-> RS
  RS --> IS["2. Idea and Script: 3 scripts per ref, n-gram 6-word check"]
  IS --> CG1{"3. Compliance Gate (pre)"}
  CG1 -->|"green"| RK["4. Recording Kit: teleprompter queue + Discord ping"]
  CG1 -->|"yellow"| AP["Chris one-tap approve in Discord"]
  AP --> RK
  CG1 -->|"red"| BL["Blocked or rewritten"]
  RK --> HU(["Chris records (the human job)"])
  HU --> QG["5. Quality Gate: Gemini + ffmpeg on Modal"]
  QG -->|"REDO with timestamped fixes"| RK
  QG --> CG2{"6. Compliance Gate (post) on actual transcript"}
  CG2 -->|"spoken hard-block: REDO"| RK
  CG2 --> ED["7. Editing: Submagic API or Modal ffmpeg"]
  ED --> PB["8. Publishing: IG + YouTube direct, GHL for TikTok, FB, LinkedIn Page"]
  PB --> AN["9. Analytics and Learning"]
  AN -.->|"hook weights, what to swipe next"| RS
  AN -.->|"rubric recalibration"| QG

Agent cards

Three compliance lanes

4. Quality Gate scorecard

100 points; compliance is pass/fail on top. Thresholds are recommendations to be recalibrated after 25 published posts against IG skip rate and watch ratio.

Decision rules

5. Tool comparison

Sort by any column; search filters all columns. Prices are from the report and research notes; most third-party prices are unverified.

6. Data model

Supabase (Postgres + pgvector). Swipe tables are new; the rest is from the report.

Cloudflare layout

7. Cadence & costs

Volume targets (weeks 1 to 4)

Pillar mix: 40% consumer protection education, 25% agent career/recruiting (yellow), 15% retirement/annuity concepts (yellow, no numbers), 20% behind the scenes, team, CRM. Scale to 7 Reels/week only when first-take quality-gate pass rate holds ≥70% for 3 straight weeks.

Monthly cost

8. Week plan: Wed Nov 4 to Tue Nov 10, 2026

Day 1 to 2 priority: get Swipe Intake working end to end before anything else. Click paths are "look for this menu," since vendor labels move. Checkboxes save in this browser.

Slip risks: Meta token or permission surprises, missing GHL API scopes, the YouTube private-lock claim. Each has a fallback (Zernio, or GHL for YouTube). None should move Sunday's recording session.

9. Interview

The answers that most change the design. They autosave in this browser. Export builds a markdown file to send back.

Export answers