SmithyAI
Building an end-to-end course-production system
The details
- Role
- Sole designer, builder, maintainer, and operator.
- Context
- Reforge course production; informed by my content-design work beginning in 2022–2023.
- Status
- Used in production, including the Spring 2026 AI Evals course.
- Reported result
- About 80% of generated production slides approved without edits, across more than 5,000 slides. Failed production generations were included; development/training courses were excluded. This is my tracked production figure, not an independently audited measure.
At Reforge, I was responsible for brand creative while continuing to produce course visuals and videos after we shifted to leaner teams. This overlap created a production constraint I understood firsthand: turning an expert’s script into a full course required design, review, voiceover, video editing, and QA.
To resolve this constraint, I independently designed, built, and maintained SmithyAI to connect those stages in one tool. I used SmithyAI in production for three roughly 40,000-word courses, with review and editing built into the process.
Production work I already knew
Before SmithyAI, the course workflow involved a content strategist, two designers, and a video producer. Production typically took 3-4 months. With SmithyAI, production was shaved down to 2-3 weeks and required only ~10 hours of active work on my part.
Those 10 hours included running the tool, editing slide outputs, reviewing outputs with the course creator, generating videos, making corrections and QA checks in Premiere.
I also built this tool around a practical requirement: generated work needed to be editable. An output that was difficult to correct would leave the production problem unresolved. Accordingly, both visuals, voice overs created via ElevenLabs and the creator’s voice clone, and the videos were all exported in a fully editable package.
From script → visuals → video
The workflow begins with an uploaded script. An AI agent maps the course into modules, lessons, and visual segments. A review panel lets me check that structure before another agent creates slides from a code library of templates.
Separating those stages gives me a place to correct the content mapping before it becomes a full set of visuals. It also keeps the system tied to a defined visual vocabulary rather than treating every slide as an unconstrained image-generation task.
Each module (3-4 lessons) took roughly 15 minutes to complete content mapping and generate slides. Once visuals were refined and approved, the tool took another 10 minutes to complete voice over, video, and final asset package generation.
Most corrections stayed within the tool
The tool has a slide editor that enables the user to edit copy, pick icons, switch slide templates, upload/replace images, and modify repeatable components. More complex work that needed further design moved into Figma and was re-imported to the tool.
I also built a Figma plugin that imported whole courses or individual lessons into Figma, placing visuals beside the relevant script passages for easier creator review and feedback.
After approval, the video workflow used AI VoiceOver generation (with the ability to accept pre-recorded voice overs), transcription and timing, and video assembly. Outputs include an XML file that Premiere can import as a linked, editable project. I still reviewed and corrected the finished videos in Premiere before export to ensure human approval.
A technical note on how the tool works: AI performs content mapping and slide creation within the template system. Code handles assembly and export. Humans remain responsible for content review, visual judgment, feedback, corrections, and final QA.
A real course changed the system’s inputs
AI Evals, from Reforge's Spring 2026 cohort, was the second course I produced with SmithyAI. Its creators did not have ElevenLabs voice models and recorded their own voiceovers.
To accommodate this, I modified the tool to accept uploaded recordings, transcribe them, and time the video edits to that audio. This modification took less than 2 hours and enabled the tool to handle yet another use case.
Conclusion
I formally tracked slide approval in production. 80% of slides met brand quality standards without edits in SmithyAI or Figma. Remaining work required editing or bespoke design.
By the time of the tools use in our Fall 2026 cohort, video production quality hit 95%, with very minimal edits being needed to exported lesson videos.
On average a module of 3-4 lessons took 25 minutes to generate from script to video, and cost $2-3 in AI tokens.
Credits
Course creators supplied expertise, scripts, and feedback. I designed, built, maintained, and operated SmithyAI, including its editing and review integrations. The tool used Opus 5 for content mapping/parsing, ElevenLabs for voice over generation, OpenAI for transcription, and a number of design and video libraries to handle visual composition and video editing/generation.