菜单
Run Lofn video/director pipeline, shot lists, cinematic scene prompts, and video generation planning. Do NOT use for static image-only prompts, music, QA, or evaluation.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Run the Lofn daily pipeline backed by Codex — fetch real-world facts, then generate the day's music (24 songs) and images (24→top 6) through the full Lofn pipeline with the daily rules (tri-source method, dual 3+3 constraint, emotional duality, library-only selection). Use for "daily run", "today's dailies", "run the daily pipeline", "do the daily drop", or a scheduled creative drop. Down-scalable for a quick test. Do NOT use for a single one-off competition piece (use `lofn`) or QA-only audits.
Audit, validate, and classify completed Lofn pipeline outputs against the strict quality gates, backed by Codex. Use after a lofn-music / lofn-image / lofn-video / lofn-story run, or on a suspicious partial run, to get a SHIP / REPAIR / FAIL verdict and a repair brief. Do NOT use for creative generation — this is the adversarial auditor, not the artist.
Run the full award-winning Lofn creative pipeline — art/image, music/song, video/animation, or story — backed entirely by Codex. Use when the user asks to create a contest-grade piece, a "Lofn" piece, a song/image/video/story "with the full pipeline", or names a Golden Seed/personality/panel. This is the front door; it runs Phase 0 (Golden Seed) + Phase 1 (3-panel orchestrator) then dispatches to lofn-music / lofn-image / lofn-video / lofn-story and lofn-qa. Do NOT use for casual one-line prompts the user wants answered directly without the pipeline.
Run the Lofn story/narrative pipeline (steps 00–10) backed by Codex — panel-driven prose, narrative voice, world-building, scripts. Use for stories, tales, narratives, flash fiction, scripts, or "write a story with the full pipeline". Expects a Phase-0/1 orchestrator packet from the `lofn` skill; if none exists, run `lofn` first. Do NOT use for images, music, video, or QA-only audits.
Run the Lofn video/animation pipeline (steps 00–10) backed by Codex — cinematic shot lists and motion/animation prompts (Veo 3.1 formula, synchronized audio). Use for video, film, cinematic clips, animation, animated loops, motion design, or "make a video/animation with the full pipeline". Covers BOTH live-cinematic and animation. Expects a Phase-0/1 orchestrator packet from the `lofn` skill; if none exists, run `lofn` first. Do NOT use for static images, music-only, story prose, or QA-only audits.
Run the Lofn daily pipeline backed by Claude — fetch real-world facts, then generate the day's music (24 songs) and images (24→top 6) through the full Lofn pipeline with the daily rules (tri-source method, dual 3+3 constraint, emotional duality, library-only selection). Use for "daily run", "today's dailies", "run the daily pipeline", "do the daily drop", or a scheduled creative drop. Down-scalable for a quick test. Do NOT use for a single one-off competition piece (use `lofn`) or QA-only audits.
| name | lofn-video |
| description | Run Lofn video/director pipeline, shot lists, cinematic scene prompts, and video generation planning. Do NOT use for static image-only prompts, music, QA, or evaluation. |
lofn-director writes video prompts (for Veo, Kling, Wan, etc.). It does NOT call any video generation API or tool.
Do NOT use: video_generate, veo, or any video generation tool.
Do NOT attempt to produce video files.
Your output is: .md files containing final ranked video prompts ready for rendering.
The main session or cron agent handles actual Veo/Kling API calls.
Every subagent must save artifacts to disk at EACH major pipeline step, not just at the end. If an agent times out or the session resets, partial work must already be on disk.
Write step files as you complete them. Never hold all output in memory for a single final write.
Every step (00–10) MUST receive the COMPLETE upstream context — no compression, no summarization. The orchestrator fills OVERALL_PROMPT_TEMPLATE.md once from its handoff artifacts (02_golden_seed.md, 03_orchestrator_panel_debate.md, 04_orchestrator_metaprompt.md, 06_*_handoff.md) into a single CREATIVE CONTEXT block: user request + Golden Seed + meta-prompt + personality + all 3 panels (18 voices) + 15 Special Flairs. That block is injected verbatim into the CREATIVE CONTEXT slot at the top of EVERY step file under steps/, alongside each step's prior outputs. Use the supplied Panel Ledger — do NOT invent a new panel; each Devil's Advocate / Hyper-Skeptic must dissent.
Do not use the legacy lofn-director agent for Steps 06-10. Use dedicated step agents:
| Step | Agent ID | Model |
|---|---|---|
| 00-05 | lofn-director-coordinator | deepseek/deepseek-v4-pro |
| 06 | lofn-director-step06 | openrouter/qwen/qwen3.7-max |
| 07 | lofn-director-step07 | openrouter/qwen/qwen3.7-max |
| 08 | lofn-director-step08 | openrouter/qwen/qwen3.7-max |
| 09 | lofn-director-step09 | openrouter/qwen/qwen3.7-max |
| 10 | lofn-director-step10 | deepseek/deepseek-v4-pro |
| 11 | lofn-director-step11 | openai/gpt-5.5 |
Full spec: vault/DIRECTOR_MODEL_ASSIGNMENTS.md
Every creative pipeline task MUST receive the target personality's full DNA block — not just "use the Solarpunk Bloom aesthetic."
Saying "style = X" or "use Industrial Grief mode" is insufficient — subagents default to Lofn's system-context personality. The personality block MUST include: cinematic identity, camera language preference, light/shadow philosophy, motion signature, and core directorial beliefs.
Inject this block at EVERY stage: seed, orchestrator, director coordinator, pair agents, step11. Without injection: Lofn bleed. With injection: target cinematic identity.
The correct pipeline order is: Research → Lofn-Core → Orchestrator → YOU → QA
You should be receiving an orchestrator-metaprompt.md and orchestrator-brief.md before you run.
If you do not have these files in your output directory:
core-brief.md exists (Lofn-Core ran but orchestrator skipped)Lofn-Core's job: seed + research. Orchestrator's job: panel + personality + metaprompt. Your job: steps 00-10, per-pair execution, renders.
Before this modality agent begins Step 00, it must validate a real Lofn-Core + orchestrator packet with:
python3 /data/.openclaw/workspace/scripts/validate_orchestrator_packet.py <run_dir>
The packet must use the original Lofn panel-object structure: Special Flairs, Concept Panel, Medium Panel, and Context & Marketing Panel, each with a Devil's Advocate / Hyper-Skeptic adversarial role. If validation fails, do not proceed; request/launch lofn-orchestrator work.
Every canonical step artifact must use /data/.openclaw/workspace/scripts/lofn_step_artifact_template.md and pass validate_with_retries.py before the next step.
YOU MUST ALWAYS USE THIS PATTERN. Never run all 10 steps in a single agent.
See TASK_TEMPLATE.md for the full specification. Summary:
concept_medium_pairs.json/data/.openclaw/workspace/vault/LEAN_PAIR_AGENT_INPUT_STANDARD.mdstep10_final_pair{N}.mdAfter Step 05 completes:
lofn-director-step06lofn-director-step07lofn-director-step11Why one step at a time: Running Step 06-10 in a single agent per pair was the #1 cause of timeout churn and template collapse in music.
Why: A single agent cannot faithfully execute all 10 steps without collapsing into templates. This was proven across 3 failed runs (2026-03-30). The split is the fix. It matches the original Lofn ui.py architecture exactly.
Stricter correction (2026-05-20): The split alone is not enough. Each numbered step must be a separate model call with a separate canonical artifact. Coordinator summary files and per-pair omnibus files are pipeline violations.
Lean pair-input correction (2026-05-21): Pair agents must not be handed the whole upstream packet by default. The parent/controller validates the full packet, then gives each video/animation pair agent a compact 50–100 line brief: compact Golden Seed operating excerpt, Step 05 artifact, concept_medium_pairs.json, one pair assignment excerpt, video Step 06–10 contract, tiny provenance block, and motion/model/safety blockers. This is mandatory for video and animation pair work.
PREREQUISITES:
0. Load resources/panel-of-experts.md to understand the panel of experts prompting you will use.
skills/lofn-core/SKILL.md for personality and Panel system.skills/lofn-core/PIPELINE.md for the MANDATORY execution pipeline.skills/lofn-core/OUTPUT.md for the MANDATORY artifact saving format.skills/video/TASK_TEMPLATE.md for exact output requirements.
4a. For pair-agent spawning or pair-agent execution, load /data/.openclaw/workspace/vault/LEAN_PAIR_AGENT_INPUT_STANDARD.md and use the compact pair brief; do not load the full upstream packet into ordinary pair agents.⚠️ EVERY video generation MUST follow the full pipeline: 10 steps, 3 panels, 6 pairs × 4 outputs = 24 treatments minimum. Then select the best N to return. No shortcuts.
Transform video concepts into production-ready prompts for video generation (Veo 3.1, Runway, Kling, Pika). This skill handles cinematic direction from concept to viral-ready output.
Before executing Step 00, run the random injection script to populate creative seeds:
node /root/.openclaw/workspace/scripts/random-injection.cjs video 50
This returns JSON with:
aesthetics - 50 random items from the master list (3000+ aesthetics)frames - 50 random items from video_frames.csv (1272 cinematic framing techniques)film_styles - 50 random film styles for cinematic directionInject these into Step 00's placeholders to "shake up the creative space."
| Step | File | Output |
|---|---|---|
| 00 | 00_Generate_Video_Aesthetics_And_Genres.md | 50 aesthetics, 50 emotions, 50 frames, 50 genres |
| 01 | 01_Generate_Video_Essence_And_Facets.md | Essence + cinematic style axes + 5 facets |
| 02 | 02_Generate_Video_Concepts.md | 12 distinct concepts |
| 03 | 03_Generate_Video_Artist_And_Critique.md | Director influence + critique per concept |
| 04 | 04_Generate_Video_Medium.md | Visual style assignment per concept |
| 05 | 05_Generate_Video_Refine_Medium.md | 6 best concept×style pairs |
| 06 | 06_Generate_Video_Facets.md | Scoring facets |
| 07 | 07_Generate_Video_Aspects_Traits.md | 24 treatments (6 pairs × 4 variations) |
| 08 | 08_Generate_Video_Generation.md | Full shot lists with audio direction |
| 09 | 09_Generate_Video_Artist_Refined.md | Director voice refinement |
| 10 | 10_Generate_Video_Revision_Synthesis.md | Ranking + final selection |
Read each step file and execute its instructions exactly.
| Shot | Code | Effect |
|---|---|---|
| Extreme Close-up | ECU | Intimacy, intensity |
| Close-up | CU | Emotional connection |
| Medium Shot | MS | Conversational |
| Wide Shot | WS | Context, scale |
| Establishing | ES | World-building |
| Movement | Effect | Prompt Note |
|---|---|---|
| Static | Contemplative | Easy, reliable |
| Slow pan | Reveals space | Specify direction |
| Tracking | Follows subject | Describe path |
| Orbit | Circles subject | "360° rotation" |
| Aerial/Drone | Epic scope | Works well |
| Angle | Subject Feels |
|---|---|
| Eye-level | Equal, neutral |
| Low angle | Powerful, imposing |
| High angle | Vulnerable, small |
| Dutch | Uneasy, off-balance |
Dialogue:
A woman whispers, "I remember everything."
Sound Effects:
SFX: Glass shattering, then silence.
Ambient:
Ambient: Rain on metal roof, distant traffic.
Music:
Audio: Swelling orchestral score.
| Platform | Aspect | Duration | Shots |
|---|---|---|---|
| TikTok/Reels | 9:16 | 15-60s | 3-12 |
| YouTube Shorts | 9:16 | 30-60s | 5-15 |
| Music Video | 16:9 | 3-4min | 30-60 |
| Cinematic | 16:9 | Variable | Scene-based |
Veo 3.1 Constraints:
See skills/lofn-core/OUTPUT.md for full artifact format.
Each treatment saved as individual file:
output/videos/{YYYYMMDD}_{HHMMSS}_{title_slug}_{pair}_{variation}.md
Every handoff file must contain an ICB — clearly demarcated with ⚠️ IMMUTABLE CONTINUITY BLOCK — DO NOT SUMMARIZE. Downstream agents receive it verbatim.
Required ICB Contents:
See vault/PIPELINE_CONTINUITY_STANDARD.md for full spec.
2 of 3 NO = BLOCKED. See vault/DIRECTOR_QA_DEPTH_AUDIT.md for full gate rules.
When receiving a video task:
vault/DIRECTOR_QA_DEPTH_AUDIT.md for per-step minimums and auto-fail triggersstep00_aesthetics_and_genres.md, step01_essence_and_facets.md, step02_concepts.md, step03_artist_and_critique.md, step04_medium.md, step05_refine_medium.md.lofn-director-step06 → write pair_NN_step06_facets.md; spawn lofn-director-step07 → write pair_NN_step07_shot_guides.md; spawn lofn-director-step08 → write pair_NN_step08_generation.md; spawn lofn-director-step09 → write pair_NN_step09_artist_refined.md; spawn lofn-director-step10 → write pair_NN_step10_revision_synthesis.mdlofn-director-step11 once per pair for final polish and shot density verificationlofn-qa with vault/DIRECTOR_QA_DEPTH_AUDIT.md for depth audit + Cinematic Somatic Gate + shot element verificationAfter writing each step file, run the validator BEFORE moving to the next step:
python3 /data/.openclaw/workspace/scripts/validate_step.py <step> <file>
Exit 0 = proceed. Exit 1 = read the error, fix the file, rerun until it passes.
# Full video pipeline validation sequence:
OUTDIR="/your/output/dir"
python3 /data/.openclaw/workspace/scripts/validate_step.py 00 $OUTDIR/00_aesthetics.md
python3 /data/.openclaw/workspace/scripts/validate_step.py 05 $OUTDIR/05_refined_pairs.md
python3 /data/.openclaw/workspace/scripts/validate_step.py 06 $OUTDIR/06_facets.md
python3 /data/.openclaw/workspace/scripts/validate_step.py 07 $OUTDIR/07_guides.md
python3 /data/.openclaw/workspace/scripts/validate_step.py 08 $OUTDIR/08_prompts.md
python3 /data/.openclaw/workspace/scripts/validate_step.py 09 $OUTDIR/09_refined_prompts.md
python3 /data/.openclaw/workspace/scripts/validate_step.py 10 $OUTDIR/10_final_prompts.md
The validator catches: stubs, lorem ipsum, template placeholders (Scene N, Shot N, Treatment N), identical pair sections, recycled "Similar setup" treatments, short content.
Only declare completion after all validators return exit code 0.
Steps 06-10 are NOT a single pass over all pairs simultaneously. They are 6 SEPARATE passes, one per pair, each producing 4 treatments.
If you find yourself writing one facet set "for all pairs" → STOP. You're collapsing the tree. If you find yourself with only 4 final treatments → STOP. You ran one pass instead of six. If you find yourself "selecting the best 4 across all pairs" before running per-pair → STOP. That's skipping the funnel.
This pipeline produces award-caliber work. Trust it. Execute it fully.