Analyze Creator Videos

Analyze Creator Videos

Build a creator operating system, not a decorative long report. Separate corpus facts, video-internal evidence, author claims, system inferences, external verification, and unknowns throughout.

This is the creator-research orchestrator. Route work to the canonical capability Skills instead of duplicating their judgment:

• $xiaohongshu-creator-acquisition for public identity, inventory, detail, comment, cover and media evidence;
• $creator-portfolio-annotation for every-post portfolio annotation;
• $creator-sample-selection for the shared comparison and deep set;
• $video-content-reconstruction for every selected deep video;
• $creator-research-synthesis for creator-level conclusions;
• $creator-research-evaluator for optional independent acceptance before formal Wiki promotion.

Use $compare-creators for a separate multi-creator study and $content-strategy-workbench for the user's own creation plan. Neither belongs inside a neutral Creator Dossier.

Read session-derived-infrastructure.md before creating a new run or changing the pipeline. Its gates come from the real build history that produced the reference Dashboard.

Agent-loop boundary

The normal production loop advances immutable stages and resumes only missing work. It must not change its own Prompt, Skill, Schema, or quality threshold while a creator run is active. A retry of the same revision is allowed only for a classified transient failure; a quality retry requires a changed input, contract, evidence, or code fingerprint.

Use two completion levels:

• DOSSIER_READY: full-corpus surface coverage is explicit, the registered deep set has usable Builder artifacts or explicit blockers, creator synthesis passes deterministic integrity checks, and the Dashboard exposes provisional conclusions and unknowns. Independent evaluation may be skipped.
• WIKI_READY: the current creator synthesis revision also passes independent evaluation and formal knowledge gates. Only this level may be promoted as verified Wiki knowledge.

Evaluator findings do not delete a usable Builder or Dossier result and do not trigger automatic repair. Use $session-forensics only in a separate optimization loop for retained representative traces when the runtime is slow, repetitive, drifting, or falsely passing; it audits the process and must not solve or modify the creator analysis itself.

Core model

Run two connected layers:

1. Portfolio layer: inspect every visible post at metadata/annotation level to learn the account's baseline, positioning, content system, distribution, topics, formats, publishing rhythm, and outliers.
2. Evidence layer: deeply reconstruct a bounded representative set with $video-content-reconstruction before making claims about content mechanisms.

Default deep set: high, median, arithmetic-mean-near, and low performance, 3 videos per group. Use the full corpus for quantitative context, a comparison set of 5–7 per group for pattern induction, and 3 per group for evidence-grade reconstruction. Median and mean are separate because breakout posts can pull the mean far above the typical post. If no post lies within 25% of the mean, label mean_gap and compare its nearest lower/upper neighbors instead of inventing an average tier. Do not reconstruct every video unless the user explicitly requests it and accepts the cost.

Inputs

Accept one or more:

• creator profile URL;
• normalized corpus JSON matching creator-corpus.schema.json;
• existing local creator research directory;
• optional local videos, subtitles, covers, comments, and public metrics;
• optional target account or business strategy to inform the final launch plan.

Create a new run directory. Never overwrite an existing creator artifact unless the user explicitly asks to update it. Store authenticated URLs, session state, and login information outside public outputs.

Workflow

1. Establish the analysis contract

Write analysis-contract.json and run-manifest.json before collecting or interpreting data:

• why the creator is being analyzed;
• decisions the result must support;
• available platforms and snapshot time;
• allowed public metrics;
• unavailable backend metrics;
• default full-corpus, comparison-set, and deep-set sizes;
• target account constraints, if any.

run-manifest.json is the single control plane for artifact paths, source revisions, stage status, validation status, and downstream invalidation. Never maintain a second canonical Dashboard or duplicate specification tree for the same creator run.

The Dashboard must answer why this creator is analyzed, what method was used, and which decisions can and cannot be made from the evidence.

2. Acquire or normalize the corpus

For Xiaohongshu, read acquisition-xhs.md and collection-infrastructure.md, then use $ego-browser. Reuse the authenticated hhh-01 session when available. Respect handoff and user-control rules. Do not use Chrome, Google browser automation, or hidden credential extraction.

When the user asks for broad platform discovery through Xiaohongshu's in-site AI, also read xhs-site-ai-discovery.md. Treat its answer as a candidate generator only; independently open and verify cited creators/posts before they enter the canonical corpus.

Initialize and maintain the resumable collection control plane:

node <SKILL_DIR>/scripts/init-collection.mjs \
  --out /absolute/path/run/collection-inventory.json \
  --creator-id <known-or-provisional-id> \
  --creator-name <creator-name> \
  --profile-url <profile-url> \
  --server hhh-01

Ingest sanitized profile, crawl-round, and detail observations with the bundled ingestion scripts. Never call a timeout or extraction error “end of profile.”

For local input, normalize it directly. Preserve raw snapshots separately from normalized data.

node <SKILL_DIR>/scripts/normalize-corpus.mjs \
  --input /absolute/path/raw-or-normalized.json \
  --out /absolute/path/run/creator-corpus.json

Do not silently convert missing values to zero. Record capture failures and incomplete pagination in collection-status.json.

Before portfolio analysis, build the collection gate:

node <SKILL_DIR>/scripts/build-collection-status.mjs \
  --inventory /absolute/path/run/collection-inventory.json \
  --out /absolute/path/run/collection-status.json

Require inventory_ready. Before deep reconstruction, require selected media to pass scripts/verify-media.mjs and the status to become deep_media_ready.

3. Annotate every post at portfolio depth

Read data-contract.md. For every post, add open-ended, multi-label annotations grounded in title, cover, post copy, and available video evidence:

• topic and user problem;
• content form and intent;
• hook and promised payoff;
• target audience and use context;
• proof mode;
• presenter, composition, aspect ratio, screen/UI role, text density, and result visibility;
• content mechanism and likely audience action;
• uncertainty and annotation evidence.

Do not force posts into a closed taxonomy. Add new labels when the corpus demands them. Keep unclassified when evidence is insufficient. Write a versioned portfolio-annotations.json. Every observed corpus ID must appear exactly once. A field-level unknown is a valid annotation; silently omitting a post is not. This full-corpus artifact feeds creator synthesis, while the canonical 21 selection remains the smaller comparison and deep-research index.

4. Compute the full-corpus baseline

node <SKILL_DIR>/scripts/analyze-corpus.mjs \
  --input /absolute/path/run/creator-corpus.json \
  --out /absolute/path/run/corpus-analysis.json

Calculate count, video count, mean, median, percentiles, maximum, distribution, weekday/daypart statistics, topic and format clusters, and candidate pools for high/median/low comparison.

Prefer median and distribution over mean. Display maximum alongside median and mean. Treat likes, comments, collections, and shares as public engagement signals—not exposure, retention, conversion, or follower growth.

5. Build the comparison set and choose the deep set

Read selection-method.md.

• Compare 5–7 videos per high/median/average/low group when available.
• Select 3 videos per group for deep reconstruction, reusing an overlapping median/average item only once.
• Cover distinct mechanisms, content forms, durations, eras, and evidence modes.
• Do not choose nine videos solely by ranking.
• Record selectionReason, represented mechanism, alternatives considered, and known confounds for every selected item.

If the corpus is small, shrink tiers and disclose the denominator. Never label a three-video sample “the creator's universal rule.”

6. Reconstruct selected videos

Use $video-content-reconstruction for each deep-set video. Require Builder integrity before using a reconstruction in a provisional Dossier; require independent evaluation before treating it as verified Wiki evidence. Preserve the distinction in every downstream claim.

For every selected video preserve:

• complete verbatim transcript;
• cue ↔ representative frame ↔ all overlapping shots;
• sparse and dense frame views;
• content architecture and core knowledge units;
• viewer cognitive change;
• procedural or argument relationships;
• claims, observations, inferences, and unknowns;
• full readable article that restores the video's key information.

Do not infer tool causality, product capabilities, commercial rights, performance, or success from montage adjacency.

7. Synthesize across videos

Generate creator-analysis.json matching creator-analysis.schema.json. Read analysis-method.md.

Separate observed portfolio facts, sample associations, mechanism hypotheses, optional external verification, and unknowns. Explain high vs median vs low through user value, content promise, proof strength, comprehension cost, information density, visual structure, novelty, audience fit, and CTA—not likes alone.

8. Optionally evaluate the creator run for formal promotion

Use $creator-research-evaluator when formal Wiki promotion or an explicit quality audit is requested. Keep producer and evaluator reasoning independent. Require explicit acquisition, corpus, annotation, selection, CR/DL/VE, synthesis, and projection gates. A partial section may remain visible with its blocker, but cannot be promoted to validated creator knowledge. Findings remain attached to the current revision; do not automatically repair or repeatedly rerun the candidate.

9. Build or update the Dashboard

Read dashboard-contract.md. Produce dashboard-data.json matching dashboard-data.schema.json, then run:

node <SKILL_DIR>/scripts/build-dashboard-data.mjs \
  --corpus /absolute/path/run/creator-corpus.json \
  --stats /absolute/path/run/corpus-analysis.json \
  --analysis /absolute/path/run/creator-analysis.json \
  --selection /absolute/path/run/selection.json \
  --out /absolute/path/run/dashboard-data.json

node <SKILL_DIR>/scripts/render-dashboard.mjs \
  --data /absolute/path/run/dashboard-data.json \
  --out /absolute/path/run/dashboard

Use the bundled template. When updating an existing Dashboard, preserve its URL, navigation, visual system, and working interactions; add new evidence into the existing page rather than creating a weaker parallel Dashboard.

Treat dashboard/ as generated output. Change source data, schemas, scripts, or the template; do not hand-patch a delivered Dashboard as the only source of truth.

Required surfaces:

1. analysis purpose and account positioning;
2. full-corpus baseline and public-metric boundaries;
3. topic and format portfolios with count, median, mean, and maximum likes;
4. high/median/average/low comparison and success/failure mechanisms;
5. all-post List and Gallery views;
6. selected-video evidence with transcript, frames, content architecture, and unknowns;
7. publishing rhythm and video language;
8. evidence boundaries, research method, and exact missing stages.

10. Validate

python3 <SKILL_DIR>/scripts/validate-json.py \
  --schema <SKILL_DIR>/schemas/analysis-contract.schema.json \
  --data /absolute/path/run/analysis-contract.json

python3 <SKILL_DIR>/scripts/validate-json.py \
  --schema <SKILL_DIR>/schemas/run-manifest.schema.json \
  --data /absolute/path/run/run-manifest.json

python3 <SKILL_DIR>/scripts/validate-json.py \
  --schema <SKILL_DIR>/schemas/collection-inventory.schema.json \
  --data /absolute/path/run/collection-inventory.json

python3 <SKILL_DIR>/scripts/validate-json.py \
  --schema <SKILL_DIR>/schemas/collection-status.schema.json \
  --data /absolute/path/run/collection-status.json

python3 <SKILL_DIR>/scripts/validate-json.py \
  --schema <SKILL_DIR>/schemas/creator-corpus.schema.json \
  --data /absolute/path/run/creator-corpus.json

python3 <SKILL_DIR>/scripts/validate-json.py \
  --schema <SKILL_DIR>/schemas/selection.schema.json \
  --data /absolute/path/run/selection.json

python3 <SKILL_DIR>/scripts/validate-json.py \
  --schema <SKILL_DIR>/schemas/creator-analysis.schema.json \
  --data /absolute/path/run/creator-analysis.json

python3 <SKILL_DIR>/scripts/validate-json.py \
  --schema <SKILL_DIR>/schemas/dashboard-data.schema.json \
  --data /absolute/path/run/dashboard-data.json

node <SKILL_DIR>/scripts/validate-run.mjs --run /absolute/path/run

Browser-smoke the final Dashboard on desktop and mobile. Check console errors, navigation, List/Gallery switching, selected-video switching, image loading, transcript links, overflow, and evidence-boundary visibility.

Completion gate

Do not announce DOSSIER_READY unless:

• corpus coverage and collection failures are explicit;
• full-corpus statistics are reproducible from creator-corpus.json;
• portfolio-annotations.json has exact ID parity with creator-corpus.json and preserves surface-only boundaries;
• all selected videos have selection reasons;
• every mechanism claim points to comparison or reconstruction evidence;
• no Builder-only or failed reconstruction is presented as independently validated;
• unknown backend metrics remain unknown;
• Dashboard List and Gallery both work;
• validation and browser smoke pass.

If any condition fails, report NOT_READY with the failed closure. Otherwise report DOSSIER_READY and provide the canonical run path and Dashboard URL. Report WIKI_READY only when the current creator synthesis revision also passes independent evaluation and the formal promotion gate.

Output contract

analysis-contract.json
run-manifest.json
raw/                         # immutable collection snapshots
collection-inventory.json
crawl-ledger.jsonl
creator-corpus.json
collection-status.json
corpus-analysis.json
selection.json
videos/<post-id>/            # deep reconstructions and evidence
media-verification/<post-id>.json
creator-analysis.json
dashboard-data.json
dashboard/index.html
dashboard/styles.css
dashboard/app.js
validation-report.json

Do not publish raw media, private comments, authentication data, or signed URLs to a public repository.