general-video: HyperFrames' Catch-All Video Builder (594K)

general-video is the HyperFrames skill you land on when a video brief matches no shipped workflow — and its most valuable content is the arithmetic about when not to parallelize. I read the whole SKILL.md (21,964 bytes, inside heygen-com/hyperframes, 64,315 stars, last push 2026-10-15) and scraped the skills.sh detail page the same day: 594.1K installs, weekly [18,371, 16,183, 15,346, 28,322, 57,814, 41,804, 61,758, 97,530]. The pack around it is riding that curve together — sibling media-use reads 912.1K, hyperframes-core 823.4K. What the skill actually delivers is a router that refuses to guess your intent, plus a dispatch rule that only pays for itself at scale.

It Routes on Project State, and Only the First Matching Row Counts

Four states, evaluated top-down, and the instruction is explicit: stop at the first match, never keep reading lower rows.

The brief vocabulary is deliberately tiny: flow (automation vs companion), storyboard (yes/no), and a derived mode the skill tells you never to ask the user to name. It also bans inventing synonyms for those states — an ongoing “just build it” arrives as flow: automation, storyboard: no, not as a phrase the model gets to reinterpret. Before any of that, npx hyperframes usage --json runs at the start and again at milestones; if usage is unknown, the skill wants the agent to say so rather than quote a guessed allowance.

Scaffolding is one command, and the state lives on disk rather than in the conversation:

npx hyperframes init "videos/<project>" --non-interactive --example=blank --skill=general-video

Dispatch Pays for Itself Only Past Six Scenes — They Measured It

Most “use sub-agents for speed” advice comes with no number attached. This file ships one: a film of up to ~6 short scenes builds faster inline, in the same context — measured at 5 short scenes ≈ 9 minutes inline versus ≈ 21 minutes when packetized. Fan out only when the plan exceeds that, and then the shape matters more than the trigger: 2–3 scenes per worker rather than one, and all workers spawned in a single wave, because a second wave nearly doubles the window.

The mechanics behind it are worth knowing even if you never dispatch anything:

• node <SKILL_DIR>/scripts/frame-packets.mjs writes one bounded packet per scene under .hyperframes/frame-packets/ — the scene’s storyboard block, the blueprint body, and every cited rule recipe, inlined — plus a _role.md that concatenates the worker role verbatim.
• Each worker reads only its own packets and the design truth file. It never opens STORYBOARD.md or the skill documents.
• The builder then waits on each scene’s configured src plus an adjacent <frame_id>.motion.json sidecar, and step 5 merges those sidecars into assembly.
• No delegation channel available? Fall back to serial, one packet at a time, still working from the packet alone.

The interview-ready detail: each scene needs a ## Frame N block in STORYBOARD.md with status: outline, a declared src:, and its blueprint or rules citation — even when storyboard: no. The block is the dispatch unit; the storyboard sheet is only the review surface.

Hard Cuts Are Clip Copies, and Catalog Search Needs Nothing Installed

Two habits separate a working composition from a guess.

For any trim, splice, or reorder of existing footage, the skill duplicates the same video source into multiple clip elements. Each copy gets its source range via data-media-start plus data-duration, and its authored placement via data-start. The audio stays on the clip (data-has-audio="true"), so cutting the video cuts its sound — no separate audio surgery. Keyframes are reserved for visual-property animation only: zoom, punch, pan, crop, mask, clip-path on an inner wrapper.

For anything the brief names as a look — “CRT scanlines”, “film grain”, “shimmer sweep” — the plan must first run a registry search:

npx hyperframes catalog --query "the look, in plain English" --json

That search needs nothing installed: no project, no prior add, no account. It ranks roughly 400 hosted blocks and components from any directory, so it also applies mid-build when the user asks for a look you hadn’t planned. Hand-author a look only after a search for it came back empty.

Then the gates, in order: npx hyperframes lint for fast feedback after the first HTML pass, and npx hyperframes check as the final gate — which reruns lint internally, so a redundant standalone lint right before it is waste. Music under a voice track gets carved (/hyperframes-audio → scripts/carve.mjs --comp index.html); a volume duck does not finish the mix. And the render waits for approval of the final Studio preview, in autonomous mode included — opening the preview asks nothing, and it’s the one question the autonomous path keeps.

Where general-video Sits in the HyperFrames Lanes

The skill is explicit that it does not own the pieces it depends on. Read this as a routing table, not a feature list:

Design identity resolves in a fixed order too — frame.md → design.md → DESIGN.md, first file found is brand truth. With no design spec at all, four things get settled before any composition HTML: identity grounded in the house style, one sentence naming the concept angle, an embeddable font pairing (an unbundled display font may simply not exist in cloud rendering), and a defined focal element, edge anchors, supporting detail, plus background treatment.

In short: general-video is a state machine with a measured cost model attached. The routing table kills intent guessing, the packet math stops you from burning a second wave of workers on a five-scene film, the catalog search replaces half-remembered look recipes, and check (not lint) is what actually closes a build.

Next step: run npx hyperframes skills update general-video — a successful no-op means you’re current, and the skill wants an update failure surfaced rather than papered over from memory. If you’re still choosing a framework for the video layer, the HyperFrames vs Remotion comparison lays out where each one stops being the right call.