blitzreels-clipping（闪电战剪辑）

BlitzReels Clipping

Use the /clips resource as the primary clipping path. It wraps ingest, transcription, suggestion generation, layout, captions, and QA into a single state machine driven by next_action. The agent follows next_action rather than orchestrating individual stages.

The low-level staged endpoints (suggestion apply, preview-frames, timeline/media-views, etc.) exist for manual-control scenarios but should not be used unless the /clips resource cannot produce the result.

Canonical Endpoints

User-Facing Workflow

Before creating a clip, let the user choose which segment to clip. This avoids wasting an API call on content the user doesn't want.

1. Ingest and wait for suggestions. Create the clip with selection_mode: "auto_best" and poll until source.suggestions_status = "ready". The first poll response with suggestions will include selection.alternatives.
2. Present suggestions to the user. Show the ranked list from selection.alternatives:

- title, hook, start/end timestamps, duration, score - Let the user pick one, or provide their own time range.

1. If the user picks a different suggestion than auto_best selected, call POST /clips/{clip_id}/reselect with selection_mode: "suggestion" and the chosen suggestion_id.
2. If the user provides manual timestamps, call POST /clips/{clip_id}/reselect with selection_mode: "time_range" and the absolute start_seconds + end_seconds.
3. Continue the normal poll loop — follow next_action until export.

If the user says "just pick the best one" or doesn't express a preference, skip step 2 and let auto_best proceed.

Default Path

1. Create the clip with POST /clips.
2. Poll with GET /clips/{clip_id} while next_action = "poll".
3. When suggestions are ready, present them to the user for selection (see above).
4. If next_action = "reselect", pick another candidate from selection.alternatives or switch to a manual time range, then call POST /clips/{clip_id}/reselect.
5. If next_action = "repair", call POST /clips/{clip_id}/repair once.
6. If next_action = "export", call POST /clips/{clip_id}/export.
7. Keep polling until status = "exported", status = "failed", or next_action = "stop".

The clip resource handles transcription, suggestion generation, reframe analysis, layout apply, captions, and QA internally. The agent never needs to call those endpoints directly.

Default Request

{
  "source": {
    "source_type": "youtube",
    "asset_id": null,
    "youtube_url": "https://www.youtube.com/watch?v=VIDEO_ID"
  },
  "selection": {
    "selection_mode": "auto_best",
    "suggestion_id": null,
    "start_seconds": null,
    "end_seconds": null
  },
  "target": {
    "aspect_ratio": "9:16",
    "max_duration_seconds": 75
  },
  "layout": {
    "layout_mode": "auto"
  },
  "captions": {
    "enabled": true,
    "style_id": "documentary"
  },
  "qa": {
    "qa_mode": "required"
  },
  "export": {
    "auto_export": false
  }
}

Overrides:

• If the user gives an existing asset, use source_type: "asset" with asset_id.
• If the user gives exact timestamps, use selection_mode: "time_range" with start_seconds and end_seconds.
• If the user picks a specific suggestion, use selection_mode: "suggestion" with suggestion_id.
• layout_mode: "auto" (default, evidence-based), "prefer_split" (forces dual-speaker split when evidence exists), "prefer_focus" (forces single-subject focus-cut).
• qa_mode: "required" blocks export on QA failure; "permissive" allows export with warnings.

Caption style defaults:

• documentary — default recommendation
• full-sentence — sentence captions with active-word emphasis
• single-word-instant — one-word-at-a-time captions
• Any custom theme UUID from /caption-themes — use a theme ID created via the caption themes skill

Poll Fields

Read these fields on every GET /clips/{clip_id}:

• status — current workflow state
• next_action — what to do next (poll, reselect, repair, export, stop)
• error — structured error with step, code, message, why_failed, how_to_fix, and retryable
• source.analysis_status, source.analysis_version
• source.transcript_status, source.suggestions_status
• source.canonical_duration_seconds — best-available duration (max of transcript, suggestions, and asset metadata because any single source can be wrong)
• selection.alternatives — ranked suggestion list with scores (present these to the user for selection)
• clip_window — applied start/end/duration
• layout.applied_mode, layout.primary_layout, layout.fallback_used, layout.warnings
• captions.status, captions.clip_window_aware, captions.warnings
• qa.status, qa.blocking, qa.issues, qa.preview_urls, qa.visual_debug_url
• export.status, export.download_url, export.short_download_url

Suggestion Presentation Format

When presenting suggestions to the user, format each alternative clearly:

Example:

Here are the best moments I found:

1. "Why Most Startups Fail" (score: 0.92)
   0:45 – 1:58 (73s)
   Hook: The number one reason startups fail isn't what you think

2. "The Hiring Mistake" (score: 0.85)
   3:12 – 4:28 (76s)
   Hook: We hired 10 people in 2 weeks and it nearly killed us

3. "Product-Market Fit Signal" (score: 0.78)
   7:01 – 8:15 (74s)
   Hook: The moment we knew we had product-market fit

Which one would you like to clip? Or give me a custom time range.

Decision Rules

Fast path:

• if qa.status = "passed" and next_action = "export", export

Blocked path — follow next_action, which will be one of:

• "poll" when blocking_reason is source_not_ready, suggestions_not_ready, or analysis_not_ready — the clip resource is still auto-advancing through preparation stages
• "reselect" when blocking_reason is invalid_selection (suggestion not found or time range invalid) or letterbox_rejected (planner couldn't achieve split/focus layout for podcast content — choose a different suggestion or time range)
• "repair" when blocking_reason is visual_qa_failed — QA found critical framing issues. Run one repair pass, then poll. If still blocked after repair, next_action switches to "reselect"
• "stop" when status is terminal (exported or failed)

Repair modes:

• "auto" — recommended default, re-runs reframe planning
• "prefer_split" — when the issue is speaker visibility in a podcast/interview
• "prefer_focus" — when the issue is weak single-subject framing
• "move_captions_off_faces" — when QA flagged caption-on-face overlap

Download URLs

The API returns two download URLs:

• export.download_url — presigned S3 URL (long, expires in 1 hour)
• export.short_download_url — clean redirect URL like https://www.blitzreels.com/api/v1/exports/{exportId}/download (requires auth header, 302-redirects to the presigned URL)

When presenting the download link to the user, prefer short_download_url for readability.

Reasoning Behind Key Constraints

• Follow next_action, not assumptions: the clip resource's state machine handles stage sequencing internally. Asset existence does not mean transcript, suggestions, or analysis are ready — syncClip checks each gate on every poll and auto-advances when ready.
• One repair pass, then reselect: the system caps automatic repair at one pass because repeated retries accumulate planner fallbacks (more letterbox segments) and degrade quality. After one repair, if QA still blocks, next_action switches to "reselect" so a different suggestion or time range can be tried.
• Captions are clip-window-aware by default: the clip resource applies captions scoped to the selected clip window. Source transcripts can span an entire hour — without scoping, a 90-second short would render captions from unrelated parts of the video.
• Letterbox rejection for podcasts: when layout_mode is "auto" and the planner can't achieve split/focus layouts (insufficient dual-speaker evidence), it rejects rather than silently exporting a low-quality letterboxed podcast clip.
• Duration has three sources: asset metadata (often wrong — reports full video length even after trimming), suggestion timing, and transcript timing. canonical_duration_seconds takes the max of all three to avoid clipping content.

Output Contract

Return:

• final_status: completed | blocked | failed
• clip_id
• project_id
• next_action if not completed
• blocking_reason if blocked
• selected clip window
• applied layout mode and whether fallback was used
• captions status and whether clip-window-aware
• QA status and issues
• export status
• short_download_url when exported (prefer this over download_url)

Manual-Control Fallback

When the /clips resource cannot produce the result (custom reframe plans, manual timeline edits, non-standard caption workflows), fall back to the staged endpoints. Read references/staged-endpoints.md for the full endpoint table and workflow.

References

• Read references/recovery.md only when the clip is blocked or failed and next_action guidance is insufficient.
• Read references/staged-endpoints.md only when manual-control fallback is needed.
• Read examples/youtube-to-shorts.md only when a concrete execution example is needed.