Communitygithub.com

Tera-Dark/personal-agent-skills

模型適配器,將完成的角色或插圖藍圖編譯為 Anima 可理解的英文提示詞,採用嚴謹的 Tag 區塊 + 自然語言區塊格式。處理 Anima 專用格式合約、長度預算、正向優先輸出、V1 忠實 / V2 增強模式、模型配置檔及偽影疑難排解。當使用者要求 Anima 提示詞、Anima 提示词,或提及 Anima 檢查點時使用。不負責設計——若尚無藍圖,請先透過 aesthetic-director-core 和 character-design-engine / illustration-direction 進行。

personal-agent-skills 是什麼?

Anima Prompt Compiler 是一個專門的模型適配器技能,用於將已完成設計決策的角色或插圖藍圖(blueprint)翻譯成 Anima 模型可理解的英文提示詞。它不負責設計本身,而是專注於格式轉換、長度管理、輸出模式切換及偽影排除。支援 V1 Faithful(忠實翻譯)和 V2 Enhanced(增強細節)兩種輸出模式,並採用嚴謹的 Tag 區塊 + 自然語言區塊兩段式格式,確保結構清晰且資訊完整。此技能在 Claude Code、Cursor 等支援技能架的開發工具上運作,能有效提升 Anima 提示詞的品質與穩定性,節省反覆調整的時間。

相容平台~Claude Code~Codex CLI~Cursor
npx skills add https://github.com/Tera-Dark/personal-agent-skills/tree/HEAD/02_creation/anima-prompt-compiler

在你喜歡的 AI 中提問

開啟一個已預先載入此 Agent Skill 的新對話。

說明文件

Anima Prompt Compiler

1. 定位(v2.0.0 起)

本 Skill 是適配器:把已經做完設計決定的 blueprint 翻譯成 Anima 能理解的英文。

它不再持有設計方法和品味規則。那些已經上移:

  • 品味簽名與否決清單 → personal-identity-profile
  • 九個創作動作、回饋診斷、成對示例 → aesthetic-director-core
  • 六層角色結構、服裝詞彙 → character-design-engine
  • 構圖協議、氛圍預設 → illustration-direction

輸入檢查:收到的東西有沒有一句帶動詞的命題、明確的輪廓、四層服裝、一個刺點、鎖定事實?沒有 → 這不是 blueprint,退回 creative-skill-router。使用者直接說"幫我寫個 Anima 提示詞,一個月光祭司"時,不要在這裡補設計。

保留在本 Skill 的只有 Anima 相關的東西:格式契約、長度預算、輸出模式、模型檔案、偽影排查。

2. 預設行為

  • 使用者要提示詞時只輸出提示詞,不主動生圖。
  • 英文 Tag block + Natural-language block 兩段式(§3)。
  • 預設只輸出正向提示詞。需要排除某物時優先使用正向約束翻譯("no jewelry" → 在 NL 段寫 bare neck and wrists),只有正向無法表達時才給一個簡短的獨立 Negative,並說明原因。
  • 不加 masterpiece, best quality, 8k, ultra-detailed 等空泛品質詞(禁用清單見 references/anima-model-profiles.md § 5.2)。
  • 不加權重、CFG、steps、取樣器、Clip Skip 等工作流參數,除非使用者要求。
  • 不改動 blueprint 的鎖定事實。

3. 格式契約:Tag + Natural Language

Part A — Tag block

逗號分隔的英文短標籤,承擔穩定鎖定資訊:

  • 主體數量:1girl, solo, 2girls
  • 展示類型與景別:full body, upper body, cowboy shot, white background, vertical composition
  • 外觀事實:髮色、髮型(寫形狀分布,如 long hair on the left side only, cropped short behind the right ear)、瞳色、種族特徵
  • 服裝單品:按 base → structural → signature extension → accessory 順序
  • 道具與動作:holding curved scissors, tucking hair behind ear
  • 角色身份(如果 blueprint 有)
  • 角色 / 系列觸發詞優先按 Anima 的 Danbooru 訓練集實際 token 寫法輸出,不要把角色名和作品名拆成多個 token。
  • 觸發詞中用於標識作品的括號、冒號等特殊字元會與 Anima 的權重語法衝突時,必須做跳脫。例:Danbooru 原始觸發詞為 37_(reverse:1999) 時,Anima 提示詞中寫作 37\(reverse1999\),不要寫成 37, reverse:1999、37_(reverse:1999) 或 37 (reverse1999)。

Tag block 簡潔、可掃描、不重複同義詞。設計邏輯不塞進標籤。

Part B — Natural-language block

一到兩個連貫英文段落,承擔標籤表達不了的關係:

  • 服裝層級、剪裁、內外關係(open coat over, hem showing beneath)
  • 輪廓的空間位置(behind the head, descending from one shoulder)
  • 不對稱的具體分布
  • 姿勢如何影響衣擺、袖子、頭髮、道具(blueprint 的因果鏈)
  • 密度分布(ornament concentrated around ___; ___ left plain)
  • 光源方向、衰減、材質響應
  • 刺點的精確位置(the only saturated color is ___ at ___)

NL 段不是 Tag 段的同義詞複述。它必須補充關係和層級。

模板

[comma-separated tags]

[One or two paragraphs: garment hierarchy, silhouette placement, asymmetry, pose causality, density, light, palette hierarchy, punctum location.]

4. 長度預算(規劃用,不是及格線)

任務Tag + NL 合計詞數
頭像 / 半身30–50
角色立繪 / 時裝50–80
多層展示板70–100
敘事場景插畫80–120

超預算時的刪除順序:重複形容詞 → 次要配件 → 背景枝節 → 材質細節。不刪:命題結構、主錨點、刺點、鎖定事實、因果鏈裡的關鍵詞。

5. 輸出模式

Direct Mode(使用者要求直接給)

  1. 一行設計方向
  2. Tag block
  3. NL block

Standard Mode(預設)

  1. 一行 Strategy Line(來自 blueprint 的命題)
  2. V1 Faithful:忠實翻譯 blueprint,只做語言轉換、消歧、結構化排版,不做審美升級
  3. V2 Enhanced:在 V1 基礎上加入光影、材質行為、構圖細化。V2 不得改動 V1 的核心事實(髮色、瞳色、服裝件、指定姿勢);如需變體,單獨標註為 optional variant
  4. Enhancement Notes:三行以內,說明 V2 加了什麼

Deep Mode(使用者要完整分析時)

  1. 命題、鎖定資訊、構圖與服設決策(引用 blueprint)
  2. Tag + NL
  3. 簡短驗證說明

預設不輸出冗長推理和審美宣言。

6. 多主體與歧義

  • 多角色:按方位分塊([LEFT]: ... [RIGHT]: ...),每塊內顏色緊貼名詞。見 references/anima-troubleshooting.md 診斷項 04。
  • Blueprint 裡未鎖定的細節:保持開放,不擅自補全(不給未指定的角色加上耳環、腰帶、手錶)。

7. 回饋輪

生圖結果有問題時:

  • 畫面偽影(手、融色、背景搶戲、串線、白底融邊)→ references/anima-troubleshooting.md,單變數最小修復。
  • 設計問題(太平淡、太亂、不像 OC、沒人味)→ 不在這裡修。退回 evaluation-loop → aesthetic-director-core/references/feedback-diagnosis.md。改 prompt 詞治不了設計層的病。

8. 輸出前檢查

  • 輸入是 blueprint 不是需求?
  • Tag 段簡潔、具體、無同義重複?
  • NL 段寫了關係(層級 / 位置 / 因果 / 密度 / 刺點位置),不是複述 tag?
  • 鎖定事實一字未改?
  • 無空泛品質詞、無參數、無預設負面?
  • 長度在預算內,且刪的是次要項?
  • V2 沒有改 V1 的事實?
  • 遵守使用者要求的輸出格式?

References

  • references/anima-model-profiles.md — 證據分級、來源溯源、模型變體、文字編碼器相容、品質詞清單、參數參考、實驗日誌
  • references/anima-troubleshooting.md — 偽影診斷目錄、最小修復、不確定性處理
  • 測試集:tests/test-suite.md

Individual skills in this repo

This repo contains 13 individual skills — each has its own dedicated page.

Tera-Dark/personal-agent-skills

Creative direction layer that turns a vague request into one committed design idea before any blueprint or prompt is written. Runs a sequence of generative "moves" (find the obsession, plant a contradiction, pick from the tail, build causality, subtract, keep one strange thing) and produces a short Creative Brief with rejected alternatives. Use for any OC / character / illustration / fashion / key-visual request, whenever output feels generic, "AI-flavored", too plain, too busy, or when the user asks for taste, direction, 审美, 创意方向, 人味, 高级感, 不要AI味.

Tera-Dark/personal-agent-skills

Persistent identity and taste layer for Tera-Dark's creative work. Holds the aesthetic signature (restrained base, one strange thing, a hint of danger, fashion-grade garment construction, female-oriented OC/gacha sensibility), hard dislikes (watches, cyber/mech, random butterflies-roses-particles, adjective costumes), business goals and workflow style. Does not generate prompts. Load first for any creative, design, illustration, prompt or aesthetic-judgment task; also when the user says 我的风格, 我喜欢, 按我习惯, 个人偏好.

Tera-Dark/personal-agent-skills

Entry point for all creative requests in this skill hub. Classifies intent (OC/character design, illustration, fashion, NAI5 prompt, Anima prompt, image reverse analysis, prompt review, ComfyUI/LoRA/dataset), loads identity + aesthetic direction first, then hands off to the right specialist and model adapter. Use whenever a request involves 设计, OC, 人设, 立绘, 插画, 服装, 提示词, prompt, NAI, NovelAI, Anima, 反推, 分析图片, ComfyUI, LoRA, or when it is unclear which skill should handle a creative task.

Tera-Dark/personal-agent-skills

Turns a Creative Brief (from aesthetic-director-core) into a complete, model-agnostic character blueprint — design thesis, silhouette architecture, visual anchors, garment engineering (base / structural / signature extension / accessory system), material contrast, palette hierarchy, behavioral pose + camera, narrative residue, presentation format — then runs a subtraction pass. Use for OC, 人设, 角色设计, 服装设计, 立绘, 高定服设, 二游角色, character sheet, fashion concept. Never writes model-specific prompt syntax.

Tera-Dark/personal-agent-skills

Model adapter that compiles a finished character or illustration blueprint into a natural-language image prompt for models that are not Anima or NAI5 — Midjourney, DALL-E / GPT Image, Google Imagen / Gemini image, Flux, SDXL-style checkpoints, or an unnamed target. Keeps parameters out of the prompt unless the target officially supports them, and labels every model-specific claim with an evidence level. Use when the user names any other image model, says 通用提示词, MJ, Midjourney, DALL-E, Imagen, Flux, SD, 或没说用什么模型. Does not design — if no blueprint exists, route through aesthetic-director-core first.

Tera-Dark/personal-agent-skills

Turns a Creative Brief into an authored, model-agnostic illustration blueprint — the image thesis, captured moment, visual motif, environment-character relationship, camera and framing, scale contrast, negative space, physical light sources, density map, narrative residue, and one coherent surreal detail. Use for 插画, 氛围图, 竖屏, 半留白, 印象风, 故事感, key visual, poster, scene illustration, cinematic composition, or whenever a character needs to be placed into a designed moment rather than displayed on a plate. Never writes model-specific prompt syntax.

Tera-Dark/personal-agent-skills

Model adapter that compiles a finished character or illustration blueprint into NovelAI V5 community-format prompts — weighted artist stack, global style layer, scene base, char1/char2 blocks, source#/target#/mutual# interaction tags, and targeted negative steering with weight::tag:: syntax. Use when the user asks for NAI5, NovelAI, NAI提示词, tag prompt, Danbooru-style prompt. Does not design — if no blueprint exists, route through aesthetic-director-core and character-design-engine / illustration-direction first.

Tera-Dark/personal-agent-skills

Decomposes a reference image into transferable design language — silhouette and mass distribution, motif grammar, garment construction, material contrast, palette hierarchy, pose logic, negative space and presentation format — instead of a tag dump. Output feeds aesthetic-director-core (for original work) or a model adapter (for faithful reproduction). Use for 反推, 分析图片, 提取提示词, 还原风格, 参考这张图, reverse prompt, image analysis.

Tera-Dark/personal-agent-skills

Reviews an existing image-generation prompt by separating design intent from model syntax — what it is trying to say, which design layer is weak (thesis, silhouette, garment structure, causality, density, punctum), what is noise, and what a rewrite should change. Use for 优化提示词, 这个 prompt 哪里有问题, prompt review, 分析这段提示词, why does this prompt not work.

Tera-Dark/personal-agent-skills

Technical placeholder for ComfyUI workflow organization, node troubleshooting, metadata and LoRA/model integration. Use when the user mentions ComfyUI, workflow json, nodes, 工作流. Does not handle design or prompt creativity.

Tera-Dark/personal-agent-skills

Technical placeholder for image dataset organization, caption/tag management and training preparation. Use when the user mentions dataset, 数据集, 打标, captions, 训练集. Does not handle visual design decisions.

Tera-Dark/personal-agent-skills

Technical placeholder for LoRA training pipelines: dataset prep, caption organization, training configuration, evaluation. Use when the user mentions LoRA, 训练, fine-tune, 炼丹. Does not decide character concepts or artistic direction.

Tera-Dark/personal-agent-skills

Evaluates a generated image or compiled prompt against the original Creative Brief and blueprint, using a fixed rubric (identity preservation, outfit binding, spatial clarity, unrequested additions, V1/V2 consistency, output contract) plus a design-layer read (thesis, silhouette, causality, density, punctum, one strange thing). Diagnoses which layer failed and routes the fix to that layer with a single-variable change. Use on any feedback round: 太平淡, 太乱, 不像, 没人味, 这版可以, 评价一下, 哪里不对, compare versions.

相關技能