Skip to content

Videoform: the plan and its tasks (VF1-VF13) - #207

Merged
shinyobjectz merged 1 commit into
shinyobjectz:mainfrom
wackyza:styles-plan
Nov 20, 2026
Merged

shinyobjectz merged 1 commit into
shinyobjectz:mainfrom
wackyza:styles-plan

Conversation

@wackyza

@wackyza wackyza commented Nov 20, 2026

Copy link
Copy Markdown
Collaborator

A plan for videoforms: the kind of video a person makes again and again (a listing walkthrough, a recipe short, a podcast clip), read once from five to ten of their finished videos, then followed and checked on every cut, fast. Docs only: docs/videoform.md (the plan) and docs/videoform.feature (its rask tasks, with their file plans; rask check passes).

The decisions in it

  • A videoform is a kind of fingerprint. fingerprint.ts gains an optional videoform section (context, references, vocabulary, template, rules), so it's found, inherited, taught and shared as any fingerprint (FP14).
  • It flows into what's built, not beside it:
    • its ranges go into its own profile (FP6e), with measures added by name;
    • its rules go into a controller's rule fields, which raea check PROJECT already runs (DR4, DR8);
    • its structure goes into a NAME.script.ts template;
    • its references use brief.ts's references shape.
  • Read once, never needed again. References come from the local drive, or are streamed from a URL with nothing downloaded. Platform pages are resolved by yt-dlp, which raea fetches, pins and keeps current. Only per-shot embeddings and measures are kept: about 1.5 MB for ten one-minute references. Following and checking never open a reference.
  • Shipped models only in the product. The domain vocabulary is written by the person's own AI from a context line. The rights prompt appears once per account.
  • Holds up when the references don't agree: too few, mixed kinds (split into separate videoforms), traits that vary, or outliers. Encoder upgrades keep the stored rows' pins until the references are refreshed.
  • Judged on the person's own test sets: leave one out, telling sets apart, edit to it, and the person's blind verdicts, kept as evidence. An automatic LLM judge for the bench is deferred.

Where the code goes (from the task plans)

New crates raea-io-resolve, raea-app-videoform and raea-sys-videoform. The fingerprint crates are each within 150 lines of their limit, so they grow only by a field (+20), a profile measure map (+30) and a hook in raea check (+15).

Before building

The first proof that matters is VF4 with VF12: on at least five of the person's sets, the numbers alone must tell videoforms apart.

🤖 Generated with Claude Code

A videoform is the kind of video a person makes again and again, read once from five to ten of their finished videos
(from the drive, or streamed from a URL with nothing downloaded) into embeddings and measures, then followed and checked
on every cut without going back to the references. It's a kind of fingerprint, and flows into what raea has: its
ranges into its profile (FP6e), its rules into a controller's rule fields (DR4, DR8), its structure into a script
template, its references into brief.ts's shape. What ships runs on raea's own models; the domain vocabulary is written
by the person's own AI. docs/videoform.md is the plan (what's built on, what a videoform keeps, the resolver, when it
doesn't hold together, judging, the build order); docs/videoform.feature its rask tasks, with their plans.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@shinyobjectz

Copy link
Copy Markdown
Owner

Thanks. The plan is thorough, and making a videoform a kind of fingerprint, rather than a parallel system, fits how the rest is built. It's docs only and rask check passes, so there's nothing here to break. Whether to take it on, and in what order, is the maintainer's call. These are the points I've flagged for that decision:

  • Fetching references from platform pages with yt-dlp (fetched, pinned and kept current by raea). Even streaming with nothing kept, this touches each platform's terms. It deserves its own decision, separate from the local-drive path, which has no such question.
  • The scale: three new crates, plus growth in the fingerprint crates, which are each within 150 lines of their limit.
  • VF4 with VF12 as the first proof: the numbers alone telling videoforms apart on five of the person's sets. That's a good gate. I'd keep it as the condition for starting the rest.

@shinyobjectz

Copy link
Copy Markdown
Owner

Decided: merging this as the plan for videoforms. It's docs only, rask check passes, and it already carries the safeguards I asked about:

  • a rights prompt once per account;
  • login-only and private links refused;
  • only measurements and vectors kept, never the video.

The build order is the one the plan sets: VF4 with VF12 first, from the local drive, as the gate. Resolving platform pages (VF1b, yt-dlp) is built only once that gate passes on the person's own sets. Thanks.

@shinyobjectz
shinyobjectz merged commit db53845 into shinyobjectz:main Nov 20, 2026
4 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants