The Problem
AI can generate impressive images and videos, but the cost adds up fast — especially when you are iterating. A single image generation call costs pennies, but a full content pipeline (research, writing, images, character consistency, animation) can cost dollars per piece. For someone producing content regularly, these costs compound.
The real issue is not the per-call price. It is that most pipelines treat every step the same — sending everything through expensive models regardless of whether the task requires that level of capability. I wanted to design a system where the architecture itself controls cost, not just the individual model choices.
Total per 6-panel strip: ~$0.33 — 90% of creative work validated before any expensive API runs
Approach: Four Cost Layers
The entire system is designed around one principle: do the maximum amount of work before the prompt ever reaches an expensive API. Work flows through four cost layers, from free to expensive.
Everything starts with a story capture step where I articulate an experience into a structured brief. Then strategy and format selection, then writing and editing the storyboard in plain markdown. The storyboard is parsed deterministically using pattern matching — no API call at all. By the time any model is involved, the creative direction is already locked.
Research, script parsing, blog drafting, and editing. All text-based work. Uses GPT-4o for quality at pennies per call. The script is structured enough that the model's job is refinement, not invention.
Character reference sheets and comic panel generation. Real reference photos are fed in to maintain approximately 98% character consistency across panels — which avoids costly regeneration. Every panel is cached on disk; if it already exists, the system skips it. A 6-panel comic strip costs about $0.29–0.33 total for images.
This is the most expensive step and it only runs after everything else is reviewed and approved. Text-to-speech narration, then video animation of each approved panel. If the video generation fails, the system automatically falls back to creating a static video from the panel image — zero extra cost.
Key Findings
UX design is a cost control mechanism.
Human checkpoints at three stages (story brief, style confirmation, panel review) prevent expensive API calls on unapproved work. By the time the most expensive API runs, 90% of the creative work is already validated. This is not just workflow design — it is financial architecture expressed through user experience.
Story-first enforcement prevents waste.
An always-on rule enforces that no content gets created without a story brief tracing back to a real experience. The AI agent cannot start generating expensive content on its own — the human must initiate and approve at every stage. This was a deliberate design constraint, not a technical limitation.
Graceful degradation is more valuable than perfect output.
When the video generation model fails (which happens), the system automatically produces a static video from the panel image using a free local tool. The user gets output either way. Designing for failure is as important as designing for success.
Skills as Workflow Templates
The system uses "skills" — structured workflow templates that define how each type of content gets created. For example, the comic scripter skill requires that a story brief already exists before it will start. It walks through choosing a comic pattern (reveal, comparison, escalation, dialogue, before/after), writing the script with scene descriptions and dialogue for each panel, defining a character design brief for visual consistency, and only then handing off to the generation step.
The pipeline supports stepwise execution — I can run panels only first, review them, fix any that need changes, and only then trigger the video step. I can also process a single scene at a time instead of all at once. This granular control means I never pay for work I have not approved.
Constraints & Reflections
This system is designed for a single user (me) with specific creative preferences. The cost architecture assumes a particular content cadence and style. A team-based version would need collaborative review steps, role-based approvals, and shared caches.
Character consistency at 98% means approximately 1 in 50 panels needs regeneration. For short comic strips this is manageable, but for longer-form content (20+ panels), the compounding probability of at least one inconsistency becomes near-certain. A more robust approach would use a dedicated character model or LoRA fine-tuning.
I have not yet measured the time cost of the human review steps. While they save money on API calls, they add latency to the pipeline. For some content types, the time cost of review may outweigh the financial savings — a tradeoff I have not formally quantified.