SuperAPI AI Directory

jianying-draft-video-planning

Design & creation

View repository

Updated 10-27 10:06

About this skill

Video scenario planning

Turn “what can be edited from this footage” into editing plans that users can choose and Harness can hand off. The 108 scenarios belong to one planning skill, not 108 duplicate skills, and do not mean that 108 native capabilities have been verified.

Runtime contract

  • Trigger conditions: the user provides a video goal or footage and needs to identify the video type, compare editing approaches, and select a workflow.
  • Required capability: online handshake uses capabilities; offline method planning does not call the CLI and must mark capabilities as unverified.
  • Minimum CLI version:1.6.13; execution still checks the live capability item by item.
  • Default risk:read_only。
  • Input facts: purpose, audience, publishing channel, target deliverables, footage list; detected duration and dimensions must not be fabricated.
  • Confirmation points: before scenario ambiguity, changing deliverables, adopting a paid model, uploading footage, launching a native app, or formal writing.
  • Success evidence:plan; it only indicates that candidates have been formed or the planning precheck has passed, not that editing, playback, or export has been completed.
  • Prohibited actions: do not download reference works by default, do not copy shots or copy verbatim, do not automatically call models, do not overwrite original footage, and do not treat method examples as runtime evidence.

When to use

“I have a bunch of travel footage and don’t know how to edit it” enters this skill; “full wedding ceremony or fast-cut highlight” enters candidate selection. “Move clip 3 to 2 seconds” is an already specified atomic edit; hand it off directly to jianying-draft-edit without redoing scenario planning.

Capability boundaries

Can provide sourced editing methods, scenario selection, and capability gaps. Real footage is required to complete shot judgments; cannot infer picture content from file names, and cannot treat keyword matching as visual analysis or automatic video generation.

Open catalog compatibility

Legacy plan-video continues to use the frozen v1 catalog. Plugins that support workflow-plan explicitly read v2 catalog and contract. The scale of v2 is determined by the actual number of entries in the catalog scenes , and no quantity, family count, or Recipe count is used as a scaling upper limit; the current catalog contains 108 seeds. Each scenario additionally declares skills、required_capabilities、beats and editorial_intent, used to route to specific skills and CLI capabilities. The 12-dimension vocabulary does not mean every scenario has been fully annotated; currently only domain mappings have been migrated, and other unknown conditions must be clarified. Three representative workflow knowledge vectorsrespectively freeze the footage gaps, purposes, Recipes, outputs, and acceptance boundaries of Vlog, multi-version course, and multi-deliverable wedding workflows; they are formal knowledge, not runtime success records. Quality profile catalogand itsstrict contract provide itemized thresholds, hard gates, evidence types, allowed rework action categories, and manual judgment boundaries for these three workflows. Read the corresponding profile only when a workflow has been selected and a quality contract needs to be compiled; it does not grant execution, approval, model invocation, or export permissions, and profile thresholds must not be treated as evidence that a real final video has passed. Auxiliary production catalogProvide five Recipes: software tutorial screen recording, README installation tutorial, intelligent zoom product demo, Web data visualization intro, and film commentary. The directory contains provider_required、 capability_gated and plan_only are stop conditions, not evidence of an installed Provider or of having passed real CapCut. When no case matches, new strategies may be proposed, but scene IDs, execution permissions, or media timecodes must not be fabricated. When an old plugin lacks a new entry point, keep the old behavior; do not copy the development directory to overwrite its skill lock.

Step 1–6 — Selection order

  1. First identify the channel, purpose, and delivery format. If the channel is unknown, custom may be used first, but do not invent platform specifications; before export, fill in dimensions, duration, and real-time platform limits.
  2. Open the scene index. If the scene ID is explicit, select it directly; if "wedding video" covers multiple examples such as quick cut, full-length, and vows, give 2–3 differentiated candidates and explain the differences.
  3. Read only the selected example; compare its minimum assets, story segments, key editing methods, and unique acceptance points. If assets are missing, list reshoots or adjustment plans; do not directly fabricate Plan timecodes.
  4. According to Recipe organize the steps. Structured directory is the legacy routing source of truth; channel profile is planning advice, not an official hard platform limit. When screen recording, intelligent zoom, Web VFX, README tutorial, or film commentary is matched, then read the auxiliary production directory; do not rewrite its steps as free-form commands.
  5. Get the installed binary's jianying capabilities --json, and check according to the selected delivery and optional features; a plan without voiceover is not required to have TTS.
  6. Output the selected scene, Recipe, asset gaps, manual steps, required skills/capabilities, confirmation points, and delivery evidence boundaries. See the process at workflow。

Capability handoff

Planning belongs to this skill, task status, approval, and retry belong to the plugin Harness, and media atomic operations belong to the Rust CLI. When entering a deterministic task, hand off to jianying-draft-schema; installation: npx skills add full-aigc-skills/jianying-skills --skill jianying-draft-schema. Execution hand off to jianying-draft-edit; installation: npx skills add full-aigc-skills/jianying-skills --skill jianying-draft-edit. The installation commands of other target skills are written separately in each example; granular installation does not automatically include them.

Minimal Job structure only explains the existing v2 contract, and does not represent an executable edited final video containing assets. Capability mapping distinguishes atomic operations, human judgment, and native execution.

Rules — Source and quality boundaries

Source index keeps search terms, links, access dates, reading scope, and popularity boundaries for each scene. Having view counts does not mean the scene will be popular; if there are no values, mark popularity_unverified. Video descriptions and search excerpts must not be described as complete viewing. Steps are independently written adaptation plans, not copying media from third parties or claiming completed user cases.

Content statistics must be separated into columns: the formal directory currently contains 108 seeds;test_fixture is only used for compatibility and negative tests; the three representative workflows are knowledge contracts; compilation, playback, export, and host results can only be counted from runtime evidence and cannot be inferred from these files.

Gotchas — Failure and recovery

  • Missing assets: waiting_input, point out specific shot/audio track/fact gaps.
  • Multiple reasonable scenes: needs_selection, do not secretly choose the highest ranked one.
  • Required capability not supported: blocked_capability; do not silently lower the delivery level.
  • Execution result unknown: retain the task and input summary, verify before deciding whether to retry; the planner itself does not perform retries.
  • Native export still requires real application and artifact evidence; a proxy preview cannot pass as final acceptance.

Verification checklist

  • The selection is the user's real intended use; channel restrictions are confirmed or explicitly unknown.
  • The selected example's unique materials and acceptance conditions have entered the plan.
  • Atomic capabilities come from the binary handshake; manual/external capabilities are listed separately.
  • Material sources, likeness, commercial use, and sensitive information have been reviewed.
  • The output declares only the evidence level actually achieved.