So far this class has been about converting ideas: maps, boards, agents, review loops. Here is the uncomfortable truth underneath all of it — AI can generate images, voices, and clips in seconds, but it cannot make an empty idea worth watching. The highest-value use of AI for a creator is not generation. It is research: finding audience questions, locating source material, comparing explanations, and turning raw information into an original angle.

An empty idea is unwatchable at any production quality

Run the thought experiment. A cinematic explainer with perfect narration, licensed music, and flawless captions — about a claim nobody questions, nobody needs, and no source supports. It will be skipped in four seconds, and rightly so. Now reverse it: a plain screen recording that answers a question people actually ask, with evidence they can check. It gets saved, shared, and rewatched despite zero polish.

Production quality multiplies idea quality; it cannot replace it. A content engine fed with unresearched ideas produces expensive noise at increasing speed — exactly the spam machine Lesson 66.3 warned against. Research is the step that makes everything downstream worth doing, which is why it comes before a single generation call in the workflow below.

The vocabulary: ten terms that keep research honest

Research goes wrong mostly through blurred language. Pin these ten terms before you start:

• Research question — what you are trying to answer, stated as a question.
• Audience question — what the viewer is already asking, in their words.
• Primary source — firsthand material: papers, data, official docs, direct observation, product specs.
• Secondary source — someone else's account of primary material: articles, summaries, commentary.
• Editorial angle — your distinct take: which slice of the evidence, framed for which audience, to what end.
• Source log — where each fact came from, link included.
• Fact check — verifying an important claim against its source before publishing.
• Claim — the assertive sentence the asset will stand behind.
• Visual metaphor — the image or analogy that makes the claim graspable.
• Content brief — the decision document that turns research into production instructions.

The load-bearing distinction is primary versus secondary, and claim versus angle. An angle is your original framing; a claim is what must survive contact with the source. AI may propose angles freely. It may never invent claims — every claim traces to the packet.

The workflow: from audience question to script

Follow this sequence every time. It connects directly to tools you already own: web and data tools from Part VI retrieve material, Markdown context from Part II holds it, model assistance from Part VIII organizes it, and a human checks sources and decides what is worth saying.

audience question
  → search for primary / credible source material
  → read and collect evidence
  → ask AI to organize, compare, and surface gaps
  → check every important claim against the source
  → choose a distinct editorial angle
  → turn it into a script / visual brief

The critical step is the fourth one, and its direction matters: AI organizes and compares, the human verifies. Ask the model to summarize the sources, contrast competing explanations, list what the evidence does not cover, and propose gaps — then open the sources yourself and check every claim you intend to publish. AI accelerates thought here; it does not invent the facts that make a video sound authoritative. A confident unsourced sentence in a draft is a fabrication wearing a lab coat.

Three concrete research paths

The same workflow bends toward three different productions:

1. Faceless science/history explainer — start from credible sources (papers, museum pages, datasets), collect the mechanism or event sequence, find where popular explanations disagree, and choose an angle that resolves a real confusion. The visual metaphor comes last, serving the verified claim. 2. Product/affiliate video — start from approved product facts and real customer questions (reviews, comments, support threads). The research question is practical: what does the buyer need to believe, and what demonstration would honestly support it? Anything the product cannot do goes in the avoided-claims column, permanently. 3. Children's short — start from age-appropriate educational or theme research plus one original story idea. The question is developmental (what does a five-year-old need from this minute?) as much as factual. Lesson 66.7 builds the full series bible from this foundation.

The packet: RESEARCH-PACKET.md

Standardize the output so any draft, agent, or collaborator works from the same evidence. Use this template:

# Research Packet — [audience question]

## Sources / links (aim for 5)
## Supported facts (exact, with source each)
## Open questions (what evidence doesn't settle)
## Avoided claims (what we will NOT say)
## Audience language (phrases viewers actually use)
## Three possible angles
## Selected angle + why
## Visual opportunities
## Credit / description note (how sources appear publicly)

Two sections separate professionals from amateurs: open questions and avoided claims. Writing down what you don't know prevents the draft from quietly filling gaps with confident filler. Writing down what you refuse to claim — the exaggerated benefit, the unproven mechanism — gives reviewers something to enforce. A packet without these is a mood board, not research.

Trend research versus theft

Study what works; make your own thing. The boundary is operational, not philosophical:

• Research the structure: why does this opening hold attention, why does this format fit this topic, what job does the pacing do? Structures are learnable craft.
• Steal nothing substantive: not the claim, not the examples, not the script beats, not the visuals, not the voice. If your asset could be mistaken for a reworded copy, it is a copy.

The test from Lesson 66.1 applies in reverse: your derivative should be recognizable as your angle on shared evidence, not as their video with synonyms. Cite influences in the source log. Then close their tab and build from your packet.

Exercise: five sources, five angles, one checked selection

Pick one viewer question in your niche — something with evidence available, not pure opinion. Build the full packet with five real sources. Then, with the packet as the model's only allowed ground, ask AI for five angles: constrain the prompt so every angle must cite which packet facts it uses. Finally, source-check the central claim of your favorite angle against the original source, and select one.

Done means: a RESEARCH-PACKET.md with five sources, exact supported facts, open questions, avoided claims, five AI-proposed angles, and one selected angle with its claim verified against the source. Verify: pick any sentence in the selected angle's claim and trace it to a packet source in under a minute. Common failure: angles that sound exciting but rest on facts no source contains — the model inventing authority. Reject any angle that fails the trace, however appealing, and select from what the evidence supports.

Check your understanding

• Why is research higher-leverage than generation for a creator using AI?
• What is the difference between an angle and a claim, and which one may AI propose freely?
• Why do open questions and avoided claims belong in the packet?

In the next lesson you will put research to work in the format where it matters most visibly: faceless channels — proof that original media never needed a host on camera, only an identity, a system, and something true to say.