St. Pierre AI — Content OS

System Overview

Step 1

Topics Bank

48 topics / 3 pillars

→

Step 2

Weekly Ranker

Auto Monday 9pm

→

Step 3
Slack Brief
Top 10 + enriched

→

Step 4

Pick 3–5

Akash selects

→

Step 5

Generator

GPT-4o carousel

→

Step 6

Post + ManyChat

Comment → DM

ℹ

Zero manual research. Monday brief arrives in Slack automatically. Akash picks topics, runs batch_runner.py, produces the reels. System handles everything else.

Content Pillars

Every carousel and reel must map to one of three pillars. P2 gets the highest ranking priority — most direct revenue signal.

AI Content Creation

How to generate more leads

Ranker priority: +2

Topics: 16

AI carousels, content systems, hooks, reels, green screen, consistency

AI Lead Capture / Transfer

How to book more quotes

Ranker priority: +3 ★ highest

Topics: 16

AI Messenger, ManyChat, pre-qualification, shared leads, guarantee, booked appointments

AI Data / Analytics

How to dominate your market

Ranker priority: +2

Topics: 16

Super Pixel, $14M data moat, attribution, CPL vs CPQBC, LTV targeting, market intel

Proof Points (rotate across pillars)

Client	Result	Best for
Riley	47x ROI	P2 lead capture stories
Ashley	$717K closed revenue	P1 authority building
Valerie	$676K + 300+ appts / 90 days	P2 volume stories
Thrasher Crawl Space	$360K quarter	P3 data / scale stories
TWF	$45K in 30 days from Messenger alone	P2 speed stories

Topics Bank

Single source of truth for all carousel and reel topics. 48 rows, versioned and scored.

File: Projects/StPierre/carousel-production/topics_bank.csv

Column	Purpose	Values
`id`	Unique topic ID	1–48
`pillar_code`	Maps to P1/P2/P3	P1, P2, P3
`topic`	Full topic string fed to generator	—
`hook_angle`	Default hook framing	Platform reveal, Data reveal, etc.
`featured_result`	Proof point to anchor the carousel	Riley 47x ROI, Ashley $717K, etc.
`keyword`	ManyChat comment trigger (ALL CAPS)	AICAROUSEL, MESSENGER, etc.
`niche`	Target niche for this topic	crawl space, waterproofing, both
`status`	Lifecycle stage	pending, done, winner
`score`	Ranker output (max 14 before multiplier)	Float
`ranking_notes`	Score breakdown per dimension	proof=2\|pillar=3\|brain=1\|age=1

⚠

Status = winner applies a ×2 multiplier in the ranker — use it for topics that performed well to resurface them with variation.

Weekly Ranker

Scores all pending topics across 6 dimensions. Writes scores back to CSV. Returns top 10 + reserve 5 with diversity cap.

File: ranker.py · Run: python ranker.py or auto via weekly_brief.py

Scoring Dimensions (max 14 before multiplier)

Dimension	Points	Logic
Proof strength	+2	featured_result contains real $, %, or x figure
Pillar priority	P2=+3, P1/P3=+2	P2 highest — most direct revenue signal
Brain match	>0.75 sim=+3 · >0.5=+1	Semantic search vs video_notes — content you've studied
Winner multiplier	×2	Applied to entire score if status = 'winner'
Age bonus	+1	Pending > 14 days without being ranked
Recency penalty	−2	Same pillar produced in last 7 days

Diversity Cap

Max 2 topics per pillar in the top 10. Prevents the brief from being dominated by a single pillar.

ℹ

Run python ranker.py --dry-run to preview scores without writing to CSV. Brain semantic search is skipped in dry-run (fast mode).

Weekly Creative Brief

Auto-generated every Monday at 9pm ET. Posts to Slack #akash-notes. Pick 3–5 topics, run batch_runner.

File: weekly_brief.py · Schedule: Monday 9pm ET

Per-Topic Enrichment (GPT-4o-mini)

Field	What it generates	Token budget
Hook angle	Sharper, more specific hook than the CSV default (≤8 words)	50
Contrarian take	Challenge the obvious angle — founder who's seen 1000s of contractor accounts	120
Pattern connection	Best-matching insight from brain.video_notes via semantic search	—
Slide 1 draft	60-char max opening line, scroll-stopping, no emojis	30

✓

Output: Slack brief + outputs/creative_brief_YYYY-MM-DD.md saved locally. Zero research required Tuesday morning.

Carousel System

8-slide Instagram carousel. 1080×1350px portrait. Green screen zone lower 35% for talking head. Text/visuals upper 65%.

Generator: generator.py --topic-id N · Batch: batch_runner.py --topic-ids 1 5 12

8-Slide Structure (strict — do not deviate)

HOOK

Scroll-stop headline + sub. Curiosity gap or bold claim.

ENEMY

Name the villain. Shared leads, cold calling, form fills, no AI.

PIVOT

"AI changes this entirely for home service businesses."

PROOF

Client result with specific numbers. Real $, real ROI.

STEP 1

Mechanism step 1. Simple, visual.

STEP 2

Mechanism step 2.

STEP 3

Mechanism step 3.

CTA

Comment KEYWORD to get [thing]. Specify what they'll receive.

Copy Rules

Headline: max 7 words, punchy
Body: max 2 sentences, conversational
Every slide has one visual prompt (Higgsfield / AI image gen)
Every slide has one voiceover cue (spoken tone)
CTA keyword = ALL CAPS (feeds ManyChat trigger)
No corporate speak, no hedging, no "just wanted to check in"
No AI-zesty enthusiasm or filler warmth

Generator Output

JSON file in outputs/<id>_<slug>.json — includes slides array, caption, hashtags, production notes. Status auto-updated to done in topics_bank.csv after generation.

Reel Format — @angus.sewell Model

Canonical video style for all St. Pierre AI reels. No exceptions. If it reads like LinkedIn, rewrite it.

Format Spec

Setup

Talking head, selfie cam, home office. No music, no b-roll, no captions.

Length

60–90 seconds max.

Scroll-stop

Bold contrarian text overlay written before the script. This does the heavy lifting.

Hook

"You might be like me where you [exact pain]…" — audience identification first.

Cred frame

"I manage ads for X home service businesses — here's what I'm actually seeing…"

Body

Numbered list. Each point gets a bold title-card text overlay. Feels like a masterclass, costs zero production.

Casual sign-off. No hard CTA. "Good luck, my dudes" energy.

Voice

Casual, contractions, occasional profanity. Direct. Zero polish theater. Short punchy sentences.

Never Do

Start with "Hey guys"
End with "smash that like button" or any engagement beg
Use music or b-roll (pure talking head only)
Over-explain — trust the audience to keep up
Sound scripted — read aloud test mandatory

Infrastructure

generator.py

Single carousel generator. Calls GPT-4o via brain.ask(). Saves JSON to outputs/. Marks topic done in CSV.

Production core

batch_runner.py

Queue processor. Run with --topic-ids 1 5 12 after picking from the brief. Processes in sequence.

Production core

ranker.py

6-dimension scoring engine. Writes scores back to CSV. Brain semantic search for +1/+3 boost. Dry-run skips brain.

Ideation

weekly_brief.py

Enriches top 10 with GPT-4o-mini. Posts to Slack #akash-notes. Saves markdown locally. Monday 9pm ET auto.

Ideation · Scheduled

topics_bank.csv

48 topics. 3 pillars × 16 topics. All scored weekly. Source of truth for all carousel/reel production.

Data

yt_discovery.py

Auto YouTube search → transcript → brain.video_notes. Currently DISABLED. Manual links only — drop in youtube-summarizer.

Brain input · Disabled

AI Models Used

Task	Model	Why
Carousel copy (8 slides)	GPT-4o	Full context, best instruction-following for structured JSON
Brief enrichment (hook / contrarian / slide 1)	GPT-4o-mini	Fast, cheap — 30–120 token outputs per call
Brain pattern matching	pgvector	Semantic search over brain.video_notes

brain.video_notes — Knowledge Base

Transcripts and summaries from manually-submitted YouTube videos feed the ranker's brain match score. Richer brain = more differentiated topic scoring. Add new videos via youtube-summarizer skill in Claude.

⚠

Current brain state: 71 videos ingested. Most are AI tools / Claude automation — limited crawl space-specific content. Watchlist v2 targets contractor FB ads, Messenger lead gen, appointment funnels, and high-ticket home service content.

Distribution Funnel

Every carousel and reel feeds this funnel. The CTA on every Slide 8 drives a comment. ManyChat handles the rest.

Instagram Reel or Carousel

60–90s reel or 8-slide carousel. Slide 8 / close = "comment KEYWORD"

Comment Trigger → ManyChat Auto-DM

Viewer comments the keyword (ALL CAPS). ManyChat fires within seconds.

Lead Magnet Delivered

PDF delivered in DM. Rotates by calendar week.

W1: Ad Audit PDF · W2: GHL Client Snapshot · W6: GHL Sales Snapshot

DM Conversation → Warm Qualification

ManyChat qualifies niche + revenue. Akash closes warm convos manually.

Booked Call

Qualified lead books directly from DM conversation.

Close at $6,500

Main offer: $6,500 / 6 weeks / 40 leads / 5-job guarantee. Downsell: $4,500 / 30 leads / 3-job guarantee.

ICP: $1M+/yr revenue · gross margin 55%+ · avg job $7,500+

Content → Revenue Timeline

Phase	Weeks	Focus	Expected closes
Warming	W20–W25	Content streak, brand building, warm DM convos	0 — funnel not warm yet
Pipeline seeding	W25–W28	First qualified leads in DM pipeline	1–2 prospects
First closes	W26+	Akash closes from warm pipeline	1–2 closes/mo target
5 clients	By W52	Compound — each client = case study = more proof	5 by Dec 1, 2026

Weekly Cadence

Day	Content Action	System Action
Mon (Creative day)	Review brief, pick 3–5 topics, plan week's reels	Weekly brief auto-posts to #akash-notes at 9pm
Tue	Run batch_runner, review carousel JSON outputs	Ranker scores written to topics_bank.csv
Wed–Thu	Film reels (green screen or selfie cam)	—
Fri	Edit + schedule posts. Set ManyChat keywords.	—
Daily	Reply to warm DMs, move convos toward call	—
Thu (auto)	Drop YouTube links for transcript ingestion	yt_discovery scheduled (disabled — manual only)

ℹ

Monday is Creative day in the themed-day OS. Brief arrives Sunday night. Deep work 7:20–10:30am = content planning + batch production. Ops starts 10:30am.

Content KPIs

Streak target

90d

by Aug 1, 2026

Reels / week

3–5

from batch_runner

Lead magnets

W1, W2, W6

Target closes

by Dec 1, 2026

Warm convos

W25+

pipeline seeds

Topics banked

16 per pillar

What "Working" Looks Like

Monday brief in Slack every week without fail
3–5 carousel JSONs generated per week
Every post has a keyword CTA → ManyChat wired
At least 1 new video ingested to brain.video_notes per week
Warm DM convos increasing week-over-week by W24
Producing content without a ManyChat keyword wired (no conversion path)
Going off-pillar (P1/P2/P3 only — no lifestyle, no EyeFly ops content)
Skipping the brief — manual topic selection defeats the ranker system

Content Operating System

System Overview

Content Pillars

Proof Points (rotate across pillars)

Topics Bank

Weekly Ranker

Scoring Dimensions (max 14 before multiplier)

Diversity Cap

Weekly Creative Brief

Per-Topic Enrichment (GPT-4o-mini)

Carousel System

8-Slide Structure (strict — do not deviate)

Copy Rules

Generator Output

Reel Format — @angus.sewell Model

Format Spec

Never Do

Infrastructure

AI Models Used

brain.video_notes — Knowledge Base

Distribution Funnel

Content → Revenue Timeline

Weekly Cadence

Content KPIs

What "Working" Looks Like