7 Content Generation
The cycle of instructing in natural language and revising while watching the results spread beyond code. Images, video, audio, and the pipelines that weave these together with code. This chapter maps the landscape of content generation and analyzes what it shares with, and how it differs from, code generation.
7.1 Image Generation
Pre-draft. For assignment and status, see TOPICS.md in the repository root.
Brief. Image generation shares with code generation the same iterative structure of instructing in natural language and revising while reviewing the result, but its prompt-writing practices and how results are evaluated differ considerably. This section covers the prompt practices actually used in image generation, techniques for maintaining consistent style, and how the verification structure differs compared to code generation. As the first section of Chapter 7 (Content Generation), it extends the discussion of verification asymmetry to be covered in §8.3 to images as a concrete artifact.
Seed Questions
- In what ways do image generation prompt-writing practices (style keywords, negative prompts, seed fixing, etc.) differ from code generation prompts?
- What techniques exist for maintaining style consistency across multiple images, and how well do they actually work in practice?
- What criteria, other than whether it works or not, are used to evaluate the quality of image generation results? How is the subjectivity problem in evaluation criteria addressed?
- How has the asymmetry that code can be verified by running it while images cannot been affected prompt-writing practices themselves?
- Record as a case study your own iterative process of obtaining the desired result with an image generation tool. How many attempts were needed, and what did you change along the way to converge on the result?
Research Pointers
- Official documentation and prompt guides for major image generation models
- Practitioner guides on style control techniques (verify exact technique names and their current status)
- Document the students’ own image generation prompt iteration process and compile it as a case study
7.2 Video and Audio Generation
Pre-draft. Check the repository root’s TOPICS.md for ownership and status.
Brief. Video and audio generation is more computationally expensive than images and adds a time axis, which makes evaluation harder. This section covers the video generation model landscape as of 2025 to 2026, the current state of speech synthesis and music generation, and what criteria are used to evaluate the quality of these artifacts. Video is distinguished from §7.1 (image generation) in that it has new failure modes that images don’t have, such as frame-to-frame consistency and violations of physical laws.
Seed questions
- As of 2025 to 2026, what are the major video generation models, and how are their respective strengths and limitations reported? (Model names and release dates must be verified against primary sources.)
- How are failure modes unique to video generation (breakdown of frame-to-frame consistency, violations of physical laws, etc.) being classified?
- For speech synthesis and music generation respectively, what criteria are used to judge quality as good? Are there quantitative metrics beyond human listening judgment?
- How far have copyright disputes over video and audio generation (training data, celebrity voice cloning, etc.) progressed so far?
- Record your own experience using video or audio generation tools as a case study. What did you repeatedly adjust to get the result you wanted?
Research pointers
- Official announcements and technical blogs from major video and audio generation models (verify dates and model names against primary sources)
- Reporting on copyright and policy-related lawsuits and regulatory trends
- Record of the student’s own generation experience
7.3 Content Pipelines: Weaving Generation into Code
Pre-draft. Check ownership and status in TOPICS.md at the repository root.
Brief. Generating a single image or video is a different problem from running an entire channel by weaving planning, generation, editing, and distribution together in code. This section covers cases where content generation has been automated into a pipeline, an automated content production process that runs from script writing through video rendering to publishing. Where §7.1 and §7.2 covered individual generation techniques, this section looks at what becomes newly possible when those techniques are orchestrated in code.
Seed Questions
- What structure do actual cases of running a content channel that weaves planning, generation, editing, and distribution into a single pipeline actually have?
- How does pipeline automation manage quality instead of just increasing content volume? What differences between the output of automated channels and channels where a person intervenes every time have been reported?
- Where in the content pipeline does a person necessarily have to intervene (final approval, editing, etc.), and why has that step not been automated?
- How does the low-effort content problem created by content generation automation connect to the discussion of slop in Part 3?
- Record as a case study your own experience building, or attempting to build, a content pipeline. How far did you automate it, and where did you get stuck?
Research Pointers
- Accounts of running automated content channels (production reports on YouTube Shorts, newsletter automation, etc.)
- Examples of open-source tools or frameworks for content generation pipelines
- If this book’s author group or students run a content channel, record its pipeline structure as a case study