Vault

Media Pipeline — controlled md → video

DRAFT — scoped 2026-08-24, not yet implemented. The harness that turns a vault page into a finished video. Standing goal from project_superscalar_agents_and_media_pipeline (2026-05-29): “controlled md → video… make a project page that homes the generator list, so it’s never re-searched.” This is that page.

Links: Watching the Solver Learn (the observation-layer argument, same family), The Anchor Method, Projects SDK, career-slides skill (the existing spine — extend, don’t rebuild), cyborgdj, camelot-from-youtube, Lemonade Stand Experiment (harness-quality thesis)


The thesis this project is built on

Three independent 2026 specimens, reviewed 2026-08-24, all point the same way: output quality tracked harness quality more than model quality.

Specimen The harness What it bought
Genspark lemonade stands “Lemonade OS” — instruction layer + desktop control Two agents ran real-world businesses end to end (badly, but autonomously)
Pezzza’s PPO gladiator Live arena view + 4-signal training dashboard Two objective bugs found by watching, fixed in minutes
Stefan 3D AI — 72-hour game Unreal 5.8 native MCP + YBE add-on + Asset Hub A playable 3D platformer in 3 days, solo

Chris’s read: a good harness is the deal in creating content. This project builds ours.

Start from what exists — do not rebuild

career-slides already runs a working md → video path: vault page → Marp deck → HTML → PNG frames → ffmpegmp4. That is the spine. This project widens it, it does not replace it.

Reuse-over-rebuild applies hardest here, because the tempting move — “write a video generator” — would discard a working pipeline.

The boundary discipline (borrowed from Stefan)

The most transferable thing in the 72-hour video is not a tool, it is a refusal. He names what the AI does badly and keeps a human there:

“All models, especially Claude Code, are not that good at vision… for this task, right now, it’s not really helpful, I would say at all.” — on scene assembly and lighting, which he does by hand, along with collision meshes and collectible placement.

And the working division of labour, in one line:

“I don’t know how to do the right blueprint here, but I know that it should slide by X.”

Human supplies intent and constraint; AI supplies implementation. He also notes vagueness costs both accuracy and money — “it will spend way more tokens because it will need to do a screenshot.”

So this project must state, up front and honestly, which stages are agent-run and which are hand-run, and revise that boundary from evidence rather than optimism.

Goals

Open questions for the implementation session

Notes

Tags

methodology · ai · agents