npx skills add ...
npx skills add pluviobyte/rnskill --skill rn-replica-qc
Analyze, recreate, align, and verify reference-video motion. Use when the user asks to recreate, copy, match, align, pixel-align, or compare a video; when a HyperFrames or Remotion remake is judged not aligned; when fixed reference playback and reusable motion components must be separated; or when delivery needs full-frame runtime, render, decode, boundary, timing, and archive evidence instead of sampled screenshots alone.
npx skills add pluviobyte/rnskill --skill rn-replica-qc
Turn reference-video recreation into a red/green evidence loop and, when useful, capture the aligned motion as reusable HyperFrames or Remotion components.
Read replica-levels.md before naming or promising fidelity. A source-stream copy, an exact lossless render frame, an encoded frame-aligned MP4, a visual rebuild, and a style reference are different claims.
assets/showcases/presenton-replica-pixel-aligned-bitexact.mp4Record the reference, exact segment, renderer, fidelity label, deliverables,
component-library destination, and evidence required for approval. Use Goal mode
when available; otherwise keep alignment-report.md and patch-log.md beside
the project.
Complete when: the segment and every acceptance claim can be tested without relying on chat memory.
Probe FPS, duration, start time, pixel format, frame count, and audio presence.
For exact work, map timestamps to source frame numbers and extract by frame
index. Inclusive segment count is (end_frame - start_frame) + 1. When joining
adjacent inclusive segments, keep the earlier boundary frame and skip the first
frame of the next segment.
Use extract_halfsec_frames.py for discovery and extract_frame_range.py for
full-frame acceptance sequences.
Complete when: reference frame 1, final frame, count, and PTS mapping are explicit and reproducible.
Never present an exact-replay texture layer as a parameterized AI component.
Read alignment-workflow.md. Start with a
0.5s contact sheet, add 0.1s windows around fast motion, then repair the
earliest or largest mismatch one visible cause at a time. Sampling locates
problems; it does not approve pixel or frame-aligned delivery.
For the Presenton failure modes and the dynamic-image decode race, read presenton-lessons.md.
Complete when: the candidate reaches the declared visual state and no known earlier mismatch remains hidden behind a later fix.
For HyperFrames, npx hyperframes check . must pass before render. A successful
render log is not delivery evidence. For Remotion, use the project's equivalent
runtime and decoded-frame checks.
Complete when: all three gates are green and the report states which layer each metric proves.
When continuing a replica in segments, rerun the previous segment, render the new segment independently, skip duplicate inclusive boundary frames, and check the frame before, at, and after every join. Then probe the combined frame count and duration.
Complete when: both old and new ranges pass and no join repeats, drops, or rewinds a frame.
Register exact replay and parametric motion separately. Each entry includes:
exact-replay or parametric-motionApprove only the fidelity label proved by current evidence. If a gate is red,
say not aligned, name the first failing frame or timestamp, and continue the
replica loop.
Use a Python runtime with Pillow for frame and sprite verification. Timeline capture requires Playwright in the Node runtime.
alignment-report.md and patch-log.mdFor content-creation projects, copy final videos into 成片/ and the compact QC
evidence into the sibling 质检/ directory.