Detailed content
🍕 The Content OS pizza: balanced
In Tax OS, context accounted for ~34%; in Sales OS, tools and skills were in charge. In Content OS—the pie chart is balanced: no single layer dominates. Context and skills carry the weight, agents orchestrate, and rules & hooks — which usually come last — step in to protect the your voice from the automatic hype.
How to read: no slice crushes the others — that’s what “balanced” means. The slices teal are the foundation/action layers; the cyans (Tools and Agents) are the threads leading outward. The numbers are a reference, not a formula.
💡 The usual question
As with every OS: Which layer matters most here? e what goes stale fastest?. In the content, what goes stale quickly is the trends context (what’s trending changes every week); what’s stable is the your voice — that’s why it becomes a rule, not data to recollect.
Why learn
Because the balance tells you there’s no shortcut through a single layer here. Someone who only stacks content-generation skills without competitor context produces generic content; someone who only researches and never packages anything becomes an analyst, not a creator. The Content OS delivers when both halves work together.
Key concepts
🔬 Reverse-engineering competitors
The Content OS context starts by looking outward. Pull the top 10 videos for each competitor in your niche and reverse-engineer it: the rhythm, o angle, o default and how they tell the story of the video. Add a scan of what’s trending. This prepares your context and becomes the compendium the OS reads when it creates.
🌱 New here?
One compendium (substrate/compendium.md) is the distilled reference the OS actually reads—not the raw data. You scan dozens of videos (the raw), and an inexpensive model summarizes it into nuggets: the few sentences that matter. The harness reads the compendium, not the mountain of transcripts.
How to read: the raw videos come in on the left; the extractor isolates the patterns that matter; only the distillate (the compendium) remains for the OS to consult. You don’t paste transcripts into the context—you paste conclusions.
✓ Useful reverse engineering
- ✓Look rhythm, angle, and storyline, not just the theme.
- ✓Separate hype merchants from the legitimate ones.
- ✓Becomes a distilled compendium — reusable in every video.
✗ Shallow copy
- ✗Copy the competitor’s title without understanding why.
- ✗Dump 50 transcripts into the context and hope for the best.
- ✗Chasing trends and losing your own voice.
Why learn
Because it’s what takes your content out of guesswork. By mapping what works in the niche—and why—you make data-driven decisions instead of acting out of envy. And since everything becomes a distilled compendium, your next idea starts with context instead of a blank page.
Key concepts
🏭 Title factory (with personas)
This is the Content OS signature skill: a title factory that generates hundreds of candidates, does A/B between them and—and here’s the key—invokes teams of agents embodying the personas of the most common audience avatars (by age, by profile). Each persona gives honest feedback: “this is too clickbaity,” “you’re promising too much,” “I’ll drop off in minute two because you’re rambling.” The result is a title tested with people who will actually click.
🌱 New here?
A persona here it's an agent that pretends to be a type of viewer of yours (e.g., “skeptical beginner,” “time-strapped business owner”). A/B is comparing two options to see which one wins. Avatar is a portrait of a typical follower. You don’t ask "what’s a good title?" in a vacuum—you ask a jury that thinks like your audience.
🎯 Copy-run objective
Generate dozens of titles for a video and submit them to a jury of 3 personas from your audience, applying your anti-hype rules, then come away with 2 finalists ready for A/B testing.
Você é minha fábrica de títulos do Content OS. Tema do vídeo: <tema do vídeo>. Promessa central (a transformação que ele entrega): <o resultado que o espectador ganha>. 1. Gere 30 títulos candidatos (até 60 caracteres), variando o gatilho: curiosidade, contraste, número, medo de perder, "como eu / como fazer". 2. Monte um júri de 3 personas da minha audiência: - <persona 1: ex. "Ana, 24, iniciante, cética com hype"> - <persona 2: ex. "Bruno, 38, dono de negócio, sem tempo"> - <persona 3: ex. "Lia, 30, já me acompanha, quer profundidade"> 3. Cada persona dá nota 0 a 10 a cada título + UMA frase de porquê (ex.: "clickbait demais", "não fala comigo", "eu clicaria na hora"). 4. Aplique as minhas regras no-hype: descarte todo título com "insano", "chocante", "você não vai acreditar" ou promessa que o vídeo não cumpre. 5. Devolva os 3 finalistas em tabela (título · nota média · maior objeção) e marque os 2 que eu deveria levar para teste A/B. Não invente métricas de audiência; baseie as notas só nas personas que defini.
The parts in <...> you replace it with your data. Save the prompt as skills/fabrica-de-titulos/SKILL.md so you can call it the same thing every time.
✅ How to verify
- ✓One came out table with 3 finalists, each with an average score and the toughest objection.
- ✓No finalist contains a word banned by the no-hype rule.
- ✓The 2 marked for A/B are the highest-scoring ones with weak objections—not the most sensational.
- ✓The feedback sounds like YOUR audience (the "beginner" persona complains about jargon; the "short on time" persona complains about rambling).
Why learn
Because the title decides whether the video gets watched. Instead of guessing on your own, the factory tests dozens of options against a jury that imitates your real audience—and filters out hype before you fall in love with a title that will burn you. It’s scale with judgment, not generation in the dark.
Key concepts
🖼️ Thumbnail briefs
Thumbnails are now AI-generated, but the important part isn’t generating them—it’s the brief. The author goes through dozens of variations until finding the one that combines title and image, works on mobile, communicates the video's “TL;DR” and — ideally — shows the result/transformation before the video even opens. The thumbnail skill packages this brief so you don't have to start from scratch for every cover.
🌱 New here?
TL;DR ("too long; didn't read") is the one-line summary: the whole idea of the video at a glance. Thumbnail is the clickable cover. A brief is the structured prompt you give the image generator: what to show, the emotion, the text, the framing—so the AI doesn’t have to guess.
What a good thumbnail brief locks down
Readable on mobile
Most people watch on their phones. If the face or text disappears in a small rectangle, the cover has failed—the brief calls for high contrast and few elements.
Tell the TL;DR
The cover conveys the central idea without the title. The image and title don’t repeat the same thing; they add to each other.
Shows the transformation
The “before → after,” the promised result. Viewers see the benefit before clicking—and click for it.
💡 Practical tip
Ask the OS to generate the brief together with the finalist titles — so the cover is aligned from the start with the title the jury picked. Mismatched cover and title is the #1 mistake made by people who automate only half the process.
Why learn
Because the cover is the second filter after the title—and what most reduces the click-through rate. Packaging the brief into a skill ensures every thumbnail follows the same criteria (mobile, TL;DR, transformation), instead of relying on your inspiration that day.
Key concepts
♻️ Repurposing engines
A video isn't a video: it's raw material. The repurposing engines take the YouTube script and turn it into LinkedIn post, and the general content in a resource on Gumroad. Each engine is a skill that knows the target format—it's not “translation,” it's repackaging for the right platform while preserving your voice.
🌱 New here?
Repurpose (repurpose) is turning content into a new format for another platform. Gumroad is a store for selling PDFs, templates, and courses. LinkedIn post has its own rhythm (short hook, line breaks). A repurposing engine is the skill that handles this format conversion on its own.
The reuse chain—one source, many destinations
roteiro-youtube.md (a fonte) │ ├── repurpose-linkedin : post com gancho + quebras de linha ├── repurpose-gumroad : recurso/PDF à venda └── repurpose-newsletter : e-mail na sua voz
✓ Repurpose done right
- ✓Repackages content to fit each platform’s pace.
- ✓Keep your voice; don't turn it into robotic text.
- ✓Turns 1 production into several assets.
✗ Copy and paste
- ✗Pastes the exact same script on LinkedIn.
- ✗Losing your voice in a generic summary.
- ✗Repurpose content that didn’t even perform.
Why learn
Because it’s the biggest reach multiplier per hour worked. The expensive part—researching, scripting, recording—is already done. Repurposing engines extract 3 to 5 extra assets from every production, without making you rewrite from scratch for each platform.
Key concepts
🤖 Agents: Producer + Orchestrator
At the top, two agents. The producer takes the idea through to publication—chooses which skills to call and in what order. The orchestrator handles the writing and editing, including via Descript CLI, and wraps it all up in a care package: the finished package that goes to the human editor for final polish. Remember the course order: agents come only after context, skills, and rules are solid.
🌱 New here?
A CLI (command-line interface) is a program you operate with text commands—the agent can call the Descript CLI to edit video without clicking anything. A care package is the organized deliverable (cuts, captions, notes) that the agent prepares for the human editor to finalize.
From idea to publication. Decides which skills to call (titles, thumbnail, repurpose) and in what order.
Writing + editing via Descript CLI. Assembles the care package and delivers it to the human editor.
🔎 The order matters
These agents only work because the layers beneath them exist: the competitor compendium (context), the title factory and repurposing (skills), and the no-hype rules (guardrails). Promoting an agent before that is “hiring a manager without a team”—it has nothing to orchestrate.
Why learn
Because it’s what turns standalone skills into a production line. With a producer and orchestrator, you request "new video about X" and the OS chains together research, titles, cover art, and clips—delivering a package that the editor only has to polish. You move out of operations and into direction.
Key concepts
🚫 No-Hype Rules & the 30-40s Hook
The Content OS guardrails protect your credibility. The no-hype rule prohibits words like "insane" and any overpromising — what the video doesn't deliver, the title doesn't promise. And there's the rule of 30–40 second hook: the opening that holds the viewer’s attention, assembled in a anti-formula, without the tired cliché of every thumbnail farm.
🌱 New here?
A rule is a strong suggestion to the model (non-deterministic). A hook/hook, in the OS sense, is deterministic — always/never. Here, "30-40s hook" uses the word in the sense of video opening: the first few seconds that grab attention. Overpromising is promising more than you deliver.
✓ Under the no-hype rule
- ✓Promise only what the video delivers.
- ✓A 30–40s hook that delivers value early.
- ✓An anti-formula opening, in your voice.
✗ Banned by the rule
- ✗"Insane," "shocking," "you won’t believe it."
- ✗Promise a result the video doesn't deliver.
- ✗Wasting 2 minutes before showing anything.
⚠️ The mistake to avoid
Leave hype as a "soft rule" that you ignore when rushing the upload. If credibility is non-negotiable for your brand, raise no-hype to a deterministic hook: a reflex that blocks publication if the title contains a banned word. Deterministic where it matters.
Why learn
Because scale without guardrails destroys your brand. The more content the OS produces, the greater the risk that a sensational headline slips through. The no-hype rule and anti-formula hook ensure that speed doesn’t turn into embarrassment—you grow while keeping your audience’s trust.
Key concepts
✅ Module summary
Next module:
5.5 — Consulting OS 💼