hyperframes-animation (HeyGen HyperFrames)

Review

Codex CLI 0.166 (gpt-6-astra, low reasoning) on a dedicated Windows test PC, a fresh temporary project and temporary HOME/TMPDIR/caches per run, skill installed (treatment) or absent (control) or absent with the skill's core requirement squeezed into one sentence appended to the prompt (one-liner group), 2 runs per group per task, network off, /tmp writes blocked. The prompts never mention the skill (except the wording noted below). HyperFrames is described in one sentence in each prompt (HTML page rendered to MP4 by seeking a GSAP timeline). Tasks: a 6-second kinetic headline where three words slam in, an 8-second two-scene transition, and a 7-second count-up number with a bar. The one-liner states the animation rules in one paragraph.

Three-arm comparison

Measured with Codex CLI over 18 runs, scored by rule-based checks with no LLM judge. Total 26.1 / 30, verdict Recommended.

Verdict: Recommended (26.1 / 30). Source: heygen-com/hyperframes, skills/hyperframes-animation, skills/hyperframes-core, pinned at commit 207c8a4. Install: npx skills@latest add heygen-com/hyperframes --skill hyperframes-animation. Licence: Apache-2.0 (repo LICENSE).

What it is

The hyperframes-animation skill from heygen-com/hyperframes, tested at commit 207c8a4 (the skill is no longer at the repo HEAD, so this describes a retired version): animation guidance for HyperFrames compositions: one paused GSAP timeline registered at window.__timelines[root id], GSAP transform aliases instead of layout properties, named eases, stagger for sequences, no Math.random/Date.now/timers/rAF/getBoundingClientRect at tween time, count-up counters tweened on a proxy object, and scenes with data-start/data-duration whose total is covered by the root duration.

How we tested

Codex CLI 0.166 (gpt-6-astra, low reasoning) on a dedicated Windows test PC, a fresh temporary project and temporary HOME/TMPDIR/caches per run, skill installed (treatment) or absent (control) or absent with the skill's core requirement squeezed into one sentence appended to the prompt (one-liner group), 2 runs per group per task, network off, /tmp writes blocked. The prompts never mention the skill (except the wording noted below).

HyperFrames is described in one sentence in each prompt (HTML page rendered to MP4 by seeking a GSAP timeline). Tasks: a 6-second kinetic headline where three words slam in, an 8-second two-scene transition, and a 7-second count-up number with a bar. The one-liner states the animation rules in one paragraph.

1. Kinetic headline — Create a 6-second 1920x1080 HyperFrames (an open-source framework that renders an HTML page to MP4 by seeking a GSAP timeline frame by frame; GSAP comes from a CDN script tag) composition: a three-word headline ('Build. Ship. Repeat.') where each word slams in…
2. Scene transition — Create a 8-second 1920x1080 HyperFrames (an open-source framework that renders an HTML page to MP4 by seeking a GSAP timeline frame by frame; GSAP comes from a CDN script tag) composition with two full-screen scenes ('Problem' for 4.5s, 'Solution' for 4.5s) th…
3. Count-up stat with bar — Create a 7-second 1920x1080 HyperFrames (an open-source framework that renders an HTML page to MP4 by seeking a GSAP timeline frame by frame; GSAP comes from a CDN script tag) composition: a big number that counts up from 0 to 87% over 2.5 seconds with a horiz…

Scoring is programmatic (hidden tests / planted-issue recall / constraint checks / render+OCR where relevant), 0–5 per run. No LLM judge decided any score.

Results

The one-liner given to the third group: "In a HyperFrames composition use one paused GSAP timeline registered at window.__timelines[root id]; move things with GSAP transform aliases (x, y, scale, rotation, opacity) never width/height/top/left; pick named eases and use stagger for sequences; no Math.random/Date.now/timers/rAF/getBoundingClientRect at tween time, no repeat:-1; drive counters with a tween on a proxy object's onUpdate; scenes are timed elements with data-start/data-duration and the root data-duration must cover them."

Mean score: 3.76 without vs 4.92 with the skill (delta +1.16). Tokens per run: 131k vs 137k (1.05x); time 72s vs 70s. One-liner group: 4.52 (the skill's gain over the one-liner: +0.39; one-liner's own gain over no skill: +0.77); one-liner tokens 87k per run, so the skill costs 1.57x the one-liner.

What happened

• Control 3.76, skill 4.92, one-liner 4.52 (n=6 each). The most frequent control failures are structural: the root element lacked a complete id/width/height/duration set in 6/6 runs, and the timeline was not registered at window.__timelines in 5/6, so the HyperFrames renderer cannot find it.
• The skill passed those checks in 6/6. The one-liner registered the timeline in 2 of 6 runs only (it failed in the kinetic and count-up tasks), though it gave the root duration and attributes right.
• Kinetic headline task: both the skill and control missed per-word stagger in 2/2 runs by the stagger check (they used separate tweens per word); the one-liner used stagger in 2/2. This check is narrow, so the difference is partly about code style.
• Tokens: skill 137k, one-liner 87k, control 131k per run. The skill makes runs more reliable at the control's token cost, not cheaper, and the one-liner is 36% cheaper than either.
• The skill is large (about 133 files installed in the working folder, with rules, blueprints and techniques). Treatment runs read both hyperframes-animation and hyperframes-core, so the skill points to the core skill and part of this gain is the core contract.

Dimension scores

Limits of this test

• Static checks on the HTML/GSAP text; no frames were rendered (the hyperframes CLI, Chrome and FFmpeg were not run), so visual timing and quality are not measured.
• Tested at commit 207c8a4 because HEAD no longer contains the skill; the install command may not work against the current repository.
• The stagger check accepts only the GSAP stagger option, not equivalent per-word offsets.
• n=2 per cell.
• n = 2 per group per task: differences under about 1 point are noise.
• Rule-based scoring measures what the checks can see; it cannot judge taste or whether a result is the best possible.

Security

See security.md. Scan: 17 CRITICAL / 17 HIGH / 148 MEDIUM. Verdict: safe to run (no-network sandbox). A clean scan does not prove a skill safe.

Bottom line

With the skill the compositions meet the HyperFrames contract almost every time; without it the root and timeline registration are usually wrong. A paragraph stating the rules recovers most of that for fewer tokens. Because the skill is no longer in the vendor repo, treat the result as a record of that earlier version.