菜单
Run Lofn image/vision pipeline steps, visual prompt generation, render prompt packaging, and image competition prompt work. Do NOT use for music lyrics, QA audit, or evaluator ranking.
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-image |
| description | Run Lofn image/vision pipeline steps, visual prompt generation, render prompt packaging, and image competition prompt work. Do NOT use for music lyrics, QA audit, or evaluator ranking. |
lofn-vision writes image prompts (for FAL/Flux/GPT Image/Gemini). It does NOT call any image generation API or tool.
Do NOT use: image_generate, fal, flux, or any image generation tool.
Do NOT attempt to produce image files.
Your output is: .md files containing final ranked prompts ready for rendering.
The main session or cron agent handles actual FAL/GPT Image/Gemini 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 (step00_*.md, step01_*.md, etc.) as you complete them. Never hold all output in memory for a single final write. This was the #1 cause of lost pipeline work in music and applies equally to vision.
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_vision_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-vision agent for Steps 06-10. Use dedicated step agents:
| Step | Agent ID | Model |
|---|---|---|
| 00-05 | lofn-vision-coordinator | deepseek/deepseek-v4-pro |
| 06 | lofn-vision-step06 | openrouter/qwen/qwen3.7-max |
| 07 | lofn-vision-step07 | openrouter/qwen/qwen3.7-max |
| 08 | lofn-vision-step08 | openrouter/qwen/qwen3.7-max |
| 09 | lofn-vision-step09 | openrouter/qwen/qwen3.7-max |
| 10 | lofn-vision-step10 | deepseek/deepseek-v4-pro |
| 11 | lofn-vision-step11 | openai/gpt-5.5 |
Full spec: vault/VISION_MODEL_ASSIGNMENTS.md
Every creative pipeline task MUST receive the target personality's full DNA block — not just a name reference like "use the Solarium voice."
Saying "voice = X" or "style = Y" is insufficient — subagents default to Lofn's system-context personality. The personality block MUST include: visual identity, palette authority, material obsession, composition signature, and core aesthetic beliefs.
Inject this block at EVERY stage: seed, orchestrator, vision coordinator, pair agents, step11. Without injection: Lofn bleed. With injection: target visual 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-vision-step06lofn-vision-step07lofn-vision-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. The split per step prevents context exhaustion and gives each phase its own model-reset call.
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 image/art pair agent a compact 50–100 line brief: compact Golden Seed operating excerpt, Step 05 artifact, concept_medium_pairs.json, one pair assignment excerpt, image Step 06–10 contract, tiny provenance block, and visual/model/safety blockers. This is mandatory for art/image 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/refs/PIPELINE.md for the MANDATORY execution pipeline.skills/lofn-core/OUTPUT.md for the MANDATORY artifact saving format.skills/image/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.skills/lofn-core/GOLDEN_SEEDS_INDEX.md first (2KB), then read only the 3-4 most relevant seeds from skills/lofn-core/refs/GOLDEN_SEEDS.md using offset/limit.Pipeline step files are in skills/image/steps/ - load only the step you are currently running.
⚠️ EVERY image generation MUST follow the full pipeline: 10 steps, 3 panels, 6 pairs × 4 outputs = 24 prompts minimum. Then select the best N to return. No shortcuts.
Transform visual concepts into award-winning images. This skill executes the Lofn image pipeline from orchestrator metaprompt through final generation.
Before executing Step 00, run the random injection script to populate creative seeds:
node /root/.openclaw/workspace/scripts/random-injection.cjs image 50
This returns JSON with:
aesthetics - 50 random items from the master list (3000+ aesthetics)frames - 50 random items from frames.csv (1273 framing techniques)film_styles - 25 random film styles for cinematic languageInject these into Step 00's placeholders:
{injected_aesthetics} → comma-separated aesthetics list{injected_frames} → comma-separated frames listThis "shakes up the creative space" by ensuring every run explores different corners of the aesthetic universe.
| Step | File | Output |
|---|---|---|
| 00 | 00_Generate_Image_Aesthetics_And_Genres.md | 50 aesthetics, 50 emotions, 50 frames, 50 genres |
| 01 | 01_Generate_Image_Essence_And_Facets.md | Essence + 10 style axes + 5 facets |
| 02 | 02_Generate_Image_Concepts.md | 12 distinct concepts |
| 03 | 03_Generate_Image_Artist_And_Critique.md | Artist influence + critique per concept |
| 04 | 04_Generate_Image_Medium.md | Medium assignment per concept |
| 05 | 05_Generate_Image_Refine_Medium.md | 6 best concept×medium pairs |
| 06 | 06_Generate_Image_Facets.md | Scoring facets |
| 07 | 07_Generate_Image_Aspects_Traits.md | 24 prompts (6 pairs × 4 variations) |
| 08 | 08_Generate_Image_Generation.md | Full prompts with all required elements |
| 09 | 09_Generate_Image_Artist_Refined.md | Artist influence voice refinement |
| 10 | 10_Generate_Image_Revision_Synthesis.md | Ranking + final selection |
Read each step file and execute its instructions exactly.
| Setting | Default | Notes |
|---|---|---|
| Provider | FAL | Use fal tool |
| Model | Flux Ultra 1.1 Pro | fal-ai/flux-pro/v1.1-ultra |
| Aspect Ratio | 9:16 | Vertical/portrait (Instagram/TikTok optimized) |
| Prompts Generated | 24 | 6 pairs × 4 variations |
| Images Rendered | 12 | Top 12 after ranking |
When TARGET_RENDERER = GPT_I2 is set in the orchestrator metaprompt:
/data/.openclaw/workspace/skills/image/renderer_gpt_image2_rules.md before Step 05/data/.openclaw/workspace/vault/GPT_IMAGE2_PLAYBOOK.md for competition-grade prompt engineeringWhen TARGET_RENDERER is not set or is FLUX: use default Flux rules as documented below.
Every prompt in Steps 07-10 MUST be ≥80 words containing ALL of:
The density test: A great image prompt must be both story AND specification.
Competition-grade prompts need BOTH. The validator checks for lighting, material, and focal hierarchy terms. If it fails density, rewrite adding those elements - do not just pad with more story.
Every image MUST have an evocative title:
See skills/lofn-core/OUTPUT.md for full artifact format.
Each prompt saved as individual file with YAML frontmatter:
output/images/{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/VISION_QA_DEPTH_AUDIT.md for full gate rules.
When receiving an image task:
vault/VISION_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-vision-step06 → write pair_NN_step06_facets.md; spawn lofn-vision-step07 → write pair_NN_step07_aspects_traits.md; spawn lofn-vision-step08 → write pair_NN_step08_generation.md; spawn lofn-vision-step09 → write pair_NN_step09_artist_refined.md; spawn lofn-vision-step10 → write pair_NN_step10_revision_synthesis.mdlofn-vision-step11 once per pair for final polish and density verificationlofn-qa with vault/VISION_QA_DEPTH_AUDIT.md for depth audit + Somatic Gate + density verificationAfter writing each step file, run this validator BEFORE moving to the next step:
python3 /data/.openclaw/workspace/scripts/validate_step.py <step> <file>
If it exits with code 1 (FAIL): read the error, fix the file, rerun validation. Do NOT proceed until it passes.
# Run after each step - example for full pipeline:
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
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 prompts.
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 prompts → 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.
The per-pair execution is what creates genuine diversity. Without it, you get 4 variants of one idea instead of 24 prompts exploring 6 different creative territories.
This pipeline won competitions. Trust it. Execute it fully.