Bilibili Creator Distill

Bilibili Creator Distill Skill

When to use

Trigger this skill whenever the user wants to:

• Reverse-engineer how a Bilibili UP 主 thinks, structures arguments, or frames topics.
• Generate "what would this UP 主 say about X" outlines grounded in their public record.
• Build a reusable creator-skill package for later sessions.

Do not trigger this skill for:

• One-off video summarisation (use plain summarisation instead).
• Emulating tone without substance (this skill refuses that explicitly).
• Private / paid / leaked content.

Core principle

Distill stable, repeatable, transferable patterns from public expression. Not tone mimicry, not one-video hot takes, not ad reads. When you lack evidence, say so — do not fabricate missing layers.

Keep two outputs separate at all times:

1. How the creator talks (surface — last priority).
2. How the creator thinks (substance — first priority).

Deliverable

A creator-skill/ directory with this exact structure (see references/deliverable_template.md for file-level specs):

creator-skill/
├── SKILL.md                     # entry, 5-10 core patterns with triggers
├── soul.md                      # core drives, long-term preferences, value floor
├── limit.md                     # evidence gaps, non-transferable topics, time-bound notes
├── resource/
│   ├── content_strategy.md
│   ├── argument_patterns.md
│   ├── speech_patterns.md
│   ├── audience_dynamics.md
│   ├── key_videos_and_turning_points.md
│   └── anti_patterns.md
├── evidence/
│   ├── evidence_map.md
│   └── source_index.md
└── validation/
    ├── regression_cases.md
    └── generalization_cases.md

How to invoke scripts

This skill ships with a Python CLI under scripts/. Always prefer the CLI over ad-hoc scraping.

# One-shot end-to-end for one creator (default limit 20 videos).
python -m scripts.cli run --uid  --limit 20 --output ./output/

# Smoke test — fetch only the video index, verify fields, no downstream work.
python -m scripts.cli smoke --uid 

# Stage-by-stage (recommended for first-time distillation so the human can review each artifact).
python -m scripts.cli collect    --uid  --limit 30
python -m scripts.cli transcribe --uid  --limit 10   # optional, only when ENABLE_ASR=true
python -m scripts.cli clean      --uid 
python -m scripts.cli summarize  --uid  --backend claude-in-session
python -m scripts.cli distill    --uid 
python -m scripts.cli package    --uid  --output ./output/

Backend modes for summarize / distill:

• claude-in-session (default): CLI writes a prompt file under data/intermediate//prompts/. The running Claude session (this one) reads the prompt, reasons inline over the cleaned corpus, and writes the structured output back. Uses zero external API credits.
• anthropic-api: CLI calls the Anthropic SDK directly. Needs ANTHROPIC_API_KEY. Useful when running headless.

Required environment

• Python ≥ 3.10 and pip install -r requirements.txt once per environment.
• Optional SESSDATA in .env — unlocks full subtitle access. Without it the skill runs with reduced video coverage. See .env.example.
• ENABLE_ASR=false by default. For UP 主 whose videos have neither creator-uploaded nor AI-generated subtitles (long-form talking-head channels like 马督工), set ENABLE_ASR=true and run python -m scripts.cli transcribe --uid --limit N to fill in subtitles via yt-dlp + faster-whisper. Requires ffmpeg on PATH.

Workflow

Follow the six steps in references/workflow.md strictly, and stop at every checkpoint so the user can confirm before the next stage burns more time/tokens.

1. Build corpus — collect fetches video index + CC subtitles + pinned comments → data/raw//.
2. Clean corpus — clean strips ads, intros, outros, fillers; tags scene_type → data/intermediate//clean_corpus.jsonl.
3. Slice & summarise — summarize chunks long subtitles into slices, then rolls slice summaries up to video and topic level.
4. Extract pattern candidates — distill reads summaries and produces pattern_candidates.jsonl with evidence references.
5. Cross-video verification — promote candidates to A/B only when they recur across multiple videos and contexts. See references/evidence_rules.md.
6. Package — package renders the creator-skill/ directory from templates/creator-skill/.

Checkpoints

After each stage, post a short report to the user:

• files written (with byte counts)
• sample of 2-3 records (anonymised/truncated)
• any anomalies (missing subtitles, rate-limit backoff, schema validation failures)
• next stage that is about to run

Wait for ok / continue before moving on.

Distillation dimensions

Extract in these eight layers. Detail is in references/dimensions.md.

1. Expression DNA — opening gambit, rhythm, recurring rhetoric, transitions, closings, signature phrases.
2. Content strategy — topic sourcing, hot-take handling, selection criteria, serialisation tendencies.
3. Argument patterns — conclusion-first vs. setup-first, evidence preference (case/data/history/analogy), counter-view handling.
4. Judgment heuristics — when to commit to a verdict vs. remain reserved, risk posture, edge-case logic.
5. Audience dynamics — stance toward fans, critics, peers, platform, brand deals.
6. Honesty boundaries — where evidence is thin, where a claim is stage-bound, where extrapolation is banned.
7. Emotional value (affect) 🆕 — audience-reported emotional experience from danmaku/comments/stats. Phrased from audience side only.
8. Visual persona (persona) 🆕 — cover phash clustering, title template, duration bucket, dominant colors.

Evidence rules

See references/evidence_rules.md for the full ladder. Short form:

| Level | Meaning | Usable in | |-------|-------------------------------------------------------------------------------------------|-----------------------| | A | Recurs across multiple videos AND multiple time periods | Core SKILL.md | | B | Recurs within one theme across multiple videos | Core SKILL.md | | C | Single high-signal occurrence, worth recording | Footnote / exception | | D | Speculative, below evidence threshold | Must NOT enter output |

Output style

• Prefer accuracy over flair.
• Write judgments before tone markers.
• Each core pattern should be stateable as "when X, this creator tends to Y, because Z" — not an abstract adjective.
• Downgrade wording when evidence is thin.

Bad examples

• "他很理性。"
• "他讲得很有感染力。"

Good examples

• "遇到争议议题时，他通常先拆背景变量，再给有限结论，避免直接站队。（证据：BV1, BV2, BV3）"
• "讲复杂主题时，他倾向先给一层直觉版结论，再补结构化解释。"

Anti-patterns

You must reject each of these:

• Distilling only tone, skipping judgment.
• Generalising one viral video into a permanent rule.
• Moving ad copy into core expression DNA.
• Confusing platform persona with real personality.
• Filling evidence gaps with fabrication.
• Replacing structural rules with personality adjectives.

Honesty policy

When the user asks you to "fully role-play" as the creator, obey these:

• You may mimic stable public expression.
• You may reuse verifiable judgment frameworks.
• You must not fabricate private views, personal history, or undisclosed positions.
• You must not pretend to know topics the creator has not publicly discussed.

Recommended hedges:

• "基于公开视频内容，他更可能采用的分析路径是……"
• "公开内容里没有足够证据支持他在这个问题上的立场。"

Activation behaviour

When this skill is invoked:

1. First confirm the target creator (UID, handle, or URL) and the sample range.
2. Pull up the existing creator-skill/ if one already exists for this UID; otherwise run the pipeline.
3. Prefer matching the creator's judgment path before layering their expression style.
4. For under-evidenced queries, announce the gap before answering.
5. Phrase answers as "based on public output" rather than as the creator themself.

Reference index

• references/workflow.md — six-step pipeline with per-stage artifacts and checkpoint script
• references/dimensions.md — the eight distillation layers, detailed
• references/evidence_rules.md — A/B/C/D ladder + promotion logic
• references/distillation_prompts.md — prompt templates for slice / video / topic / pattern / packaging stages
• references/deliverable_template.md — exact file-level spec for the creator-skill/ output

Final standard

The distillation is only acceptable if the output can:

• Answer "how would he analyse this problem?" with his judgment path.
• Answer "how would he structure a video on this?" with his known scaffolding.
• Refuse gracefully when asked about topics lacking public evidence.
• Cite ≥ 2 BV-IDs for every A-level pattern.

If any of those four breaks, the result is a character card, not a skill.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: wenjiazhu1980
• Source: wenjiazhu1980/bilibili-creator-distill-skill
• License: MIT
• Homepage: https://github.com/wenjiazhu1980/bilibili-creator-distill-skill

Install and usage instructions live in the source repository linked above.