Shining Cut

Turn already-screened footage into a reference-calibrated edit while keeping every decision traceable to high-quality originals.

Start with the readiness gate

1. Inspect the supplied paths, reports, and manifests read-only. Do not assume the user ran Shining Moments.
2. Proceed only when the footage is already screened and the included material can be distinguished from rejected material. If it is unscreened, pause before reference research or editing and suggest $shining-moments. Do not silently perform culling inside this skill.
3. When the input is a Shining Moments review set, resolve its manifest and reports back to source_path and source timecodes. Treat select as primary material and review as backup. Do not silently promote memory or not_selected, and never use excluded.
4. Before speaking to the user, read references/interaction-contract.md. Keep the normal workflow to its three centralized confirmations; ask an extra question only for a genuine execution blocker.

Run the workflow

1. First confirmation — material and reference plan. Present one prefilled request covering screening status and paths, content/form, extracted keywords, optional reference enhancement, optional user reference link, and the proposed project output location.
2. Reference calibration. Read and follow references/reference-calibration.md. Automatic public-reference discovery through Google Videos and Wikimedia Commons Videos always runs. Optional enhancement defaults off and, when enabled, is exactly one of YouTube or Bilibili. Select one main reference and at most two auxiliaries using playable evidence; extract transferable rules rather than copying a work.
3. Script and storyboard blueprint. Read references/script-blueprint.md. Inventory usable video duration, recommend duration/aspect/platform, and produce a timecoded narrative, original-file mapping, representative storyboard, sound/subtitle plan, primary-grade plan, and reference rationale. Ask for the second centralized confirmation only after the complete blueprint is reviewable.
4. Rough cut. After approval, read and follow references/rough-fine-cut.md. Build a playable rough cut from mapped originals, apply the approved primary grade, and perform basic technical checks.
5. Automatic rough-cut review. Watch the rendered rough cut, inspect exact cut boundaries where needed, and create a timecoded fine-cut change list rather than rewriting the full script.
6. Third confirmation — fine-cut decision. Offer exactly: keep the rough cut, apply the recommended fine cut, or revise the fine-cut list once and then execute it. Default to retaining the rough cut's primary grade; offer secondary grading only as an explicit optional choice.
7. Fine cut, QA, and delivery. Execute the confirmed choice, run secondary grading only when the user selected it, verify the actual output, and preserve the reference record, script, storyboard, timecodes, EDL, and original mapping beside the deliverable.

Use the bundled downstream capabilities

• The complete bundle ships watch and video-use as sibling Skills at ../watch/ and ../video-use/. Prefer their registered $watch and $video-use invocations. If the host does not register sibling Skills automatically, resolve each sibling from this SKILL.md location and read its own SKILL.md; never copy or invoke an entrypoint without its companion scripts/ or helpers/ directory.
• Use watch for public video URLs or local rendered files when it can provide decoded frames and timestamped evidence. Follow its own preflight. It does not authenticate to platform accounts or accept browser cookies. Treat its output as observation evidence, not as an edit decision.
• Use video-use for actual editing and rendered-output self-evaluation. Follow its runtime setup and production-correctness rules. Pass only screened original-media mappings and the approved Shining Cut blueprint into the edit. The blueprint is the required plain-English strategy approval; do not introduce another ordinary strategy-confirmation round.
• For a YouTube or Bilibili enhancement that genuinely needs login, let the user authenticate in their own visible browser and inspect only media visible there. Never ask for, receive, store, export, or reuse a username, password, verification code, cookie, API key, or session secret. If a downstream transcript service needs an API key, ask the user to configure it outside Shining Cut or continue with evidence that does not require it.
• If either required downstream skill or its runtime is unavailable, report the concrete preflight failure and pause. Do not fabricate a tool, playback result, transcript, reference observation, or edit.

Apply editorial and grading tendencies

• For action-camera and drone footage, deprioritize or omit shots with abrupt direction changes, whip turns, sudden pans, or other jarring camera movement. Use such a shot only when the movement has clear narrative, spatial, rhythmic, or emotional meaning.
• For any explicitly included photo or video source that is flat Log or RAW material, the primary grade must normalize exposure, contrast, white balance, and color before the rough cut is delivered. After verifying the source profile and LUT compatibility, use an appropriate style LUT when helpful, such as a landscape LUT or a skin-tone-focused portrait LUT. Never hand the user a gray-looking or obviously misexposed rough cut.
• When the user requests secondary grading, use the approved reference sample and the project's subject to pursue a bolder, coherent style instead of falling back to the cautious primary-grade look. For example, Japanese travel footage can favor a Japanese-inspired palette, while family or kinship footage can favor warm, intimate tones; preserve usable exposure and natural-looking key skin tones while making the requested style clearly visible.

Preserve sources and evidence

• Default to video only. Ignore photos in mixed folders unless the user explicitly asks to add them.
• Never move, overwrite, or delete original media. Create new, versioned outputs under the confirmed project output location.
• Edit from high-quality originals, not review proxies, thumbnails, or low-resolution derivatives. Every used interval must map to an absolute original path and source time range.
• Never use page text, metadata, a search snippet, or a thumbnail as proof that a reference was watched. Record access limits honestly.
• Do not claim a rough cut, final render, playback check, or QA pass unless the corresponding artifact was actually produced and verified in this run.