Socialseal Reference Video Analysis

SocialSeal Reference Video Analysis

Overview

SocialSeal does not start content from a blank prompt. It lifts patterns from real videos that already surface for the brand's tracked searches. This skill is the evidence stage: identify the right reference videos for an opportunity, and analyze them into Video DNA you can ground a blueprint on.

Do not hand-invent "what good looks like." Pull exemplars from SocialSeal and read their actual analysis. See references/production-pipeline.md for the full pipeline and references/mcp-and-cli-usage.md for the call patterns.

Read references/creative-production-gates.md for production-bound work. The benchmark gate should let the user see representative candidates and choose the creative direction before expensive analysis or large-scale production.

When to Use

• Selecting exemplar videos for a blueprint scope (topic, competitor, tracking group, list, or manual).
• Exploring candidates with a semantic retrievalPrompt before committing.
• Extracting Video DNA for one or more specific videos or public URLs.
• Clustering reference videos to see repeated mechanisms.

Hand off the selected, analyzed exemplars to socialseal-blueprint-builder.

Inputs

Required:

• workspace id (or a configured default)
• an opportunity definition and its opportunityKey
• a scope: topic + pillarId, competitor + competitorBrandIds, tracking_group + trackingGroupId, list + listId, or manual + videoUids

Good to have:

• a retrievalPrompt describing the content you want exemplars of (topic/competitor/tracking_group scopes only)
• platformIds, region, language, timePeriod (e.g. 30d)
• pinnedVideoUids, excludedVideoUids, promotedCandidateTarget

Workflow

1. Confirm scope and workspace. Identify the opportunity and reuse its opportunityKey.
2. Preview candidates. Call vnext-blueprints-generate with previewOnly: true. Inspect the candidate and promoted-exemplar lists: each carries a score, matched keywords, sources, and metrics. Do not generate yet.
3. Offer analysis depth. Present representative candidates and let the user choose:
   • metadata/preview only (fastest; weakest visual claims)
   • focused DNA on roughly 3-5 exemplars (default for a single deliverable)
   • fuller DNA on roughly 8-12 exemplars (broader pattern confidence) Inspect the live schema/preflight and state likely time/credit trade-offs before a credit-consuming run.
4. Refine selection. Pin must-include exemplars (pinnedVideoUids), drop off-scope or off-market ones (excludedVideoUids), and set promotedCandidateTarget for how many exemplars to promote. For semantic exploration, set retrievalPrompt.
5. Analyze Video DNA. For the approved promoted exemplars, call tracked-video-extract with ensureAnalysis: true to resolve hook, content style, video structure, specific attributes, production qualities, transcript/audio/visual analysis, and signed frame/asset URLs. For a one-off public video, use extract-url with allowUntracked: true.
6. Cluster if needed. Use vnext-cluster-videos to group exemplars into repeated mechanisms when the set is large.
7. Record evidence. For every exemplar capture the human-readable citation (video title/URL, @author_handle, and the "keyword" [market, platform] it surfaced for) plus the analyzed DNA. Keep video_uid (and search_result_id where it came from a ranked row) as an internal traceability note for tool calls and blueprint joins. Remember exemplars are anecdotal creative evidence, not proof a mechanism will perform; see references/evidence-and-confidence.md.

Tool Calls (MCP-first)

Preview exemplars for a tracking-group scope:

socialseal_call_tool {
  "function": "vnext-blueprints-generate",
  "workspaceId": "<workspace-id>",
  "body": {
    "workspaceId": "<workspace-id>",
    "opportunityKey": "<opportunity-key>",
    "scopeType": "tracking_group",
    "trackingGroupId": <group-id>,
    "timePeriod": "30d",
    "promotedCandidateTarget": 12,
    "previewOnly": true
  }
}

Semantic exploration with a prompt (topic scope):

socialseal_call_tool {
  "function": "vnext-blueprints-generate",
  "workspaceId": "<workspace-id>",
  "body": {
    "workspaceId": "<workspace-id>",
    "opportunityKey": "<opportunity-key>",
    "scopeType": "topic",
    "pillarId": "<pillar-id>",
    "retrievalPrompt": "first-timer walkthroughs that answer cost and timing",
    "previewOnly": true
  }
}

Extract Video DNA for promoted exemplars:

socialseal_call_tool {
  "function": "tracked-video-extract",
  "workspaceId": "<workspace-id>",
  "body": {
    "ensureAnalysis": true,
    "includeAssets": true,
    "items": [{ "videoUid": "<video-uid>" }]
  }
}

Poll analysis status:

socialseal_get_tool_status { "id": "<video-uid>", "kind": "video_analysis", "includeResults": true }

CLI equivalents:

npx -y @socialseal/cli tools call --function vnext-blueprints-generate --workspace-id <workspace-id> --body @preview.json --pretty
npx -y @socialseal/cli video extract --video-uid <video-uid> --ensure-analysis --wait --out-dir ./video-assets --workspace-id <workspace-id>
npx -y @socialseal/cli video extract --url <public-video-url> --allow-untracked --wait --workspace-id <workspace-id>

Output

• a list of selected exemplars cited by video title/URL, @handle, and the "keyword" [market, platform] (with video_uid/search_result_id kept as a traceability note), plus score and source
• Video DNA per exemplar: hook, first frame, structure, hero/supporting shots, useful details or mood cues, production qualities, caption/CTA style, account type
• format and lens labels (aspirational vs utility/practical) with the evidence behind each
• a note on which exemplars are analyzed vs metadata-only
• approved benchmark direction and analysis-depth choice

Do / Don't

Do:

• preview before generating; refine with pins and exclusions
• get the user's benchmark/analysis-depth choice before credit-heavy DNA work
• inspect actual analysis/frames before claiming a visual or format pattern
• cite exemplars by title/URL and @handle; keep video_uid/search_result_id as a traceability note for the blueprint
• separate owned, creator, media, affiliate, and competitor exemplars

Don't:

• hand-pick "good examples" from outside SocialSeal when scoped exemplars exist
• infer a pattern from a single example, or present an exemplar as proof a hook will work
• make platform-age claims when published_at is blank (see references/socialseal-data-contract.md)
• copy hooks verbatim; capture the mechanism

Troubleshooting

• Preview returns missing_data: the scope has no qualifying evidence. Widen the keyword/scope, change timePeriod, or pick a different scope before generating.
• Exemplars lack analysis: tracked-video-extract with ensureAnalysis: true, then poll video_analysis status.
• retrievalPrompt rejected: it is not allowed for manual or list scope; use a topic/competitor/tracking_group scope.
• Too many viral outliers: lower promotedCandidateTarget and pin representative exemplars instead.

Verification Checklist

• Scope and opportunityKey are confirmed and reused.
• Exemplars were previewed and refined, not blindly generated.
• Each exemplar has a video_uid and source recorded.
• Video DNA is from real analysis or labeled metadata-only.
• Output is ready to ground socialseal-blueprint-builder.