jianying-narration

JianYing Narration

Convert long-form spoken content into auditable keep/drop decisions and isolated drafts. Cut points come from audio facts and semantics, not manual estimation.

When to use

• The user explicitly requests the CapCut capability corresponding to this skill's title, and accepts handling it through the unified Rust jianying CLI.
• The user needs an auditable plan, execution, or verification evidence, rather than merely a conceptual answer.
• When requirements are not yet clear, first hand off to jianying-use for routing; this skill should not be used to handle other editors, general video knowledge Q&A, or speculative execution lacking necessary materials.

Capability boundaries

• ✅ Can do: Generate a deterministic plan based on real paths and probe results; execute domain operations when capability is supported ; output the evidence level actually achieved.
• ⚠️ Requires conditions: Writes must use an isolated copy; costs, external services, and native application actions require precise approval; playback or export conclusions must have corresponding runtime evidence.
• ❌ Out of scope: Does not automatically install or upgrade the CLI, does not generate Python draft scripts, does not call external headless checkout, does not overwrite source drafts, and does not fabricate resource IDs, media facts, or success status.

Step 1 — Establish a factual baseline

Record the user's goal, absolute paths, host platform, target deliverables, and permitted side effects; if information is insufficient, first provide a read-only diagnostic plan and list the missing facts item by item.

Step 2 — Handshake capabilities

Run version, doctor, and capabilities checks. A satisfied version does not mean the capability is available;partial、external_dependency or a missing state must enter a degraded path.

Step 3 — Validate inputs

Probe media, drafts, configuration, or task state; reject non-existent paths, out-of-range times, unknown schemas, mismatched hashes, and approvals that cannot be bound to a target.

Step 4 — Form a deterministic plan

Write the operations, parameters, output paths, risk levels, confirmation points, and expected evidence into the plan. Order multiple tasks as "read-only facts → reversible writes → external/native actions."

Step 5 — Execute or stop before the gate

Execute only steps that are supported and authorized. When output is ambiguous, times out, or partially succeeds, retain the task ID, audit log, and artifacts; do not automatically retry.

Step 6 — Verify and deliver

Run structural validation, then append cold reopen, playback, or native export evidence according to the target; report the actual level, failed fields, recoverable actions, and unverified items.

Rules

• The material duration, dimensions, streams, hashes, and draft version in the plan must come from tool output.
• Approval must be bound to the command, parameters, target, validity period, and budget; if any field changes, reconfirm.
• User data is processed only within local target paths and the declared external Provider boundary; unrelated credentials must not be collected, uploaded, or recorded.
• When current evidence is insufficient, use UNVERIFIED、BLOCKED or AMBIGUOUS; it is forbidden to fill in results with inference.

Gotchas

1. Treating version as capability: Always check the specific capability; do not execute directly just because the CLI version is new enough.
2. Treating structural validation as film acceptance: project verify Only proves the structural layer; playback and native export require independent evidence.
3. Directly modifying the source draft: First create an isolated copy and record the input hash; any in-place overwrite request must stop and be changed to a new output.
4. Automatic retry for ambiguous: First query the task, audit, and external artifacts; only after confirming that no side effect occurred can the user approve a retry.
5. Filling in resources or time from memory: Resource IDs come from the catalog, and time comes from probe/ASR/project data; if unavailable, block explicitly.
6. Vaguely asking for more information: First give an executable read-only plan, then explicitly list gaps such as paths, delivery levels, or authorization.

Verification checklist

• CLI identity, version, and capability have been recorded.
• All input paths, media facts, and hashes are reviewable.
• Write targets are isolated from source drafts, and confirmation points are precisely bound.
• Failure and ambiguous paths retain task ID, logs, and recovery suggestions.
• Final wording is consistent with the actual evidence level.

Execution contract

• Trigger conditions: the user requests precise spoken-word editing, removal of pauses/retakes, generation of sentence-by-sentence subtitles, or reuse of transcriptions.
• Required capabilities:media.probe、media.asr_ledger、timeline.edit。
• Minimum CLI version:1.6.0; the ASR provider must also separately declare availability status.
• Default risk:external_native_execution; pure local analysis can be downgraded to read_only.
• Input facts: source media hash/duration/audio stream, existing transcription, ASR provider identity, budget, protected words, and target pacing.
• Confirmation points: paid ASR/TTS, writing to an existing draft, playback, or native execution must each be precisely confirmed.
• Success evidence:playback; a keep-list or structural verification cannot replace listening.
• Prohibited actions: do not generate Python draft scripts, do not invoke an external headless checkout, do not manually estimate word-level cut points, silently resubmit ambiguous ASR, or delete uncertain numbers/proper nouns.

Workflow

1. Probe the material and fix the content SHA-256.
2. Preferentially reuse verified transcription artifacts with the same provider, executor identity, content hash, and normalized request.
3. Stop when no ASR capability is available; do not use an ordinary text-generation model as a transcriber.
4. Generate the keep/protect list by combining silence, repeated segments, and semantics; cut points fall in word gaps.
5. Execute timeline trim/move/remove in an isolated copy, keeping the main track continuous.
6. Import subtitles, verify structure, cold reopen, and listen to cuts, sibilance, semantics, and synchronization.

For the ledger and cut-point rules, see references/workflow.md. For isolated editing jobs, see examples/minimal-job.json。

Progressive materials

• On first execution, load normal path to obtain the end-to-end sequence.
• When encountering failure or unknown status, load failure recovery and error recovery table。
• When the user requests overreach, overwrite, or automatic retry, load boundary refusal。
• Before delivery, complete verification checklistitem by item; do not claim completion based only on a plan or the existence of files.