Project and Delivery Scope
The Media Production Toolkit supports a personal video-production workflow in which AI agents take approved stories through research, narration, media acquisition, editing, review, and delivery. By October 3, 2026, the current toolkit workflow had delivered four published videos: Kiruna, The Line, Three Gorges, and Guitar Tower, each with a final-master and posting package.
My contribution is requirements, architecture, agent direction, review, and validation. Coding agents implement the software; production agents carry out the research, media, and editing work. I approve the topic, angle, format, script, and finished film, and direct revisions when the result falls short. The project demonstrates technical delivery through an agent-assisted workflow.
The toolkit follows an earlier video-production-engine project. The four videos above belong to the current workflow; results from the precursor are separate evidence.
The Problem and Requirements
A successful render is only one part of delivering a video. The workflow also needs to preserve what was approved, explain where media came from, detect changed inputs, produce the right format, and hand over the exact reviewed film with usable publication copy.
I directed the system around these requirements:
- Keep channel-specific creative guidance and approvals under a clear human-owned authority.
- Give a running production a stable set of instructions and approved inputs.
- Preserve source, usage basis, representation, and transformation history for media.
- Support both agent editing and a human-editor handoff from the same registered media pool.
- Detect stale preparation and altered staged files before accepting a package.
- Distinguish technical checks, audiovisual review, approval, and actual publication.
These requirements shaped both the software boundaries and the operating workflow.
Architecture Decisions
One toolkit, separate channel authority
I kept the toolkit as a modular monolith. Shared sourcing techniques, specialist visual production, and integrity checks remain in one repository. A channel’s strategy, creative taste, factual standards, and approval rules stay in its own Drive workspace.
This lets another channel use the same technical capabilities without inheriting the first channel’s identity or editorial policy. Production media, caches, manifests, and private working state stay outside Git.
Freeze the running production’s inputs
At initialization, the production agent captures the required guidance into an immutable local Context snapshot with source revisions and hashes. The running production uses that snapshot rather than repeatedly consulting mutable instructions.
Approved script and narration files also become part of prepared input identity. A changed request, narration, format, route, or production root requires fresh route preparation before packaging.
Hash verification has a specific limit: it can detect changes to captured content, but it cannot establish that the operator captured every required source. Completeness and approval status still require review.
Assign each component a concrete job
| Component | Responsibility |
|---|---|
| Channel workspace | Creative and editorial guidance, approvals, and durable deliverables |
| Production agent | Coordinate the stages, inspect results, and route defects back for repair |
| Sourcing and editing skills | Operating technique for finding, assessing, and assembling media |
| Remotion / Three.js | Specialist explanatory visuals when deterministic or procedural work helps |
| Handoff package | Context integrity, registration, provenance, input freshness, packaging, and technical verification |
| Tesseract or human editor | Assemble the film through the selected editing route |
| Human approval | Accept the story, script, and complete film before publication |
The integrity package does not make creative decisions. Tesseract remains an external editing runtime; the toolkit uses its supported interfaces rather than maintaining a second timeline engine.
Directing the Agents
The production agent develops a topic, angle, and format for my approval, then coordinates research and a joined script/visual plan. The workflow calls for a separate editorial read before the script reaches me. Factual accuracy and a plausible media plan are checked alongside whether the story actually holds together.
Once I approve the script, agents generate and review narration, translate story beats into explicit media needs, and acquire, derive, or generate suitable assets. Those choices preserve whether a picture is direct evidence or an explanatory representation. A generated illustration does not become documentary observation merely because it looks convincing.
Viable media is inspected, registered, and curated before editing. The agent route stages cleared registered files into local edit inputs, uses Tesseract for assembly, and draws on specialist visuals where useful. The human-editor route receives a verified package and brief. It is a supported choice when the requested treatment or finishing calls for it; it can be selected without first attempting an agent edit.
The same ownership model applies throughout: agents perform implementation and production work, while I set requirements, inspect the result, resolve tradeoffs, and make approval decisions.
QA and Recovery
Repair the input boundary before continuing
New external media introduced during editing returns through acquisition, registration, and curation. A changed media request triggers fresh preparation before repackaging. The workflow therefore has an explicit recovery path when an edit discovers a missing asset or sound requirement.
Repository regression fixtures cover changed Context bytes, changed package inputs, and stale preparation. They also cover format identity: a landscape profile rejects a portrait export, and an explicit Short profile requires its exact dimensions. These checks verify defined integrity and format conditions.
Inspect the rendered film
Agents review actual renders for narration-picture alignment, framing, text collisions, weak holds, joins, and sound problems. Revision direction identifies the affected moment and intended correction. The workflow records visual playback and listening separately, including partial or unavailable review.
When preview and export disagree, the documented recovery method compares the same affected interval in the source, preview, bounded export, and final output before another full render. This narrows the defect and preserves the original inputs and editable project during repair.
Keep technical acceptance and approval distinct
Final verification checks staged-input integrity, project readability, and output properties. Its passing result does not prove creative quality, attest every packaged editing resource, or prove that a particular MP4 came from a particular project.
I review the complete film and approve it or request revisions. Factual, script, narration, media, and rights problems return to their owning stage. Approval then permits promotion of the exact reviewed master bytes into the deliverable package.
Delivered Results and Limits
The current workflow delivered four published videos with final-master and posting packages. The public outputs are Kiruna Church Move, The Line, Three Gorges Dam, and Guitar Tower.
The delivery unit includes the approved film and posting information. Current Short packaging also provides cover options and destination-specific copy; an upload package and an actual published post remain distinct states.
Publication reconciliation registers an actual public post after matching it to the approved package. It does not infer publication from a successful export, a Drive upload, or an approval.
The result is an inspectable path from approved inputs to delivered and published work, supported by reusable production and integrity tooling. Four deliveries with linked public outputs establish that the workflow has been used beyond a software demo. They do not establish an autonomous production service, a repeatable long-form creative standard, or a particular audience-growth outcome.
This case study draws on the toolkit’s architecture, operating skills, regression fixtures, approved delivery records, and publication records reviewed for the October 3 snapshot. Private production state and source records remain private.