Anima Prompt Compiler
1. Positioning (since v2.0.0)
This Skill is an adapter: it translates a finished design blueprint into English that Anima can understand.
It no longer holds design methods or taste rules. Those have been moved upstream:
- Taste signature and rejection list →
personal-identity-profile - Nine creative actions, feedback diagnosis, paired examples →
aesthetic-director-core - Six-layer character structure, clothing vocabulary →
character-design-engine - Composition protocol, atmosphere presets →
illustration-direction
Input check: Does the received content have a proposition with a verb, a clear silhouette, four layers of clothing, a punctum, and locked facts? No → this is not a blueprint, return to creative-skill-router. When the user directly says "write me an Anima prompt for a moon priestess," do not supplement the design here.
Only Anima-related things remain in this Skill: format contract, length budgets, output modes, model profiles, and artifact troubleshooting.
2. Default Behavior
- When the user asks for a prompt, output only the prompt; do not proactively generate images.
- Output in two-part format: English Tag block + Natural-language block (§3).
- By default, output only positive prompts. For exclusions, prioritize positive constraint translation ("no jewelry" → in NL block write
bare neck and wrists); only when positive cannot express it, provide a short standalone Negative and explain why. - Do not add vague quality tags like
masterpiece, best quality, 8k, ultra-detailed(prohibited list inreferences/anima-model-profiles.md§ 5.2). - Do not add workflow parameters like weights, CFG, steps, sampler, Clip Skip unless the user requests them.
- Do not alter the blueprint's locked facts.
3. Format Contract: Tag + Natural Language
Part A — Tag block
Comma-separated short English tags, responsible for stable locking information:
- Subject count:
1girl,solo,2girls - Display type and shot:
full body,upper body,cowboy shot,white background,vertical composition - Appearance facts: hair color, hairstyle (write shape distribution, e.g.,
long hair on the left side only, cropped short behind the right ear), eye color, ethnic features - Clothing items: in base → structural → signature extension → accessory order
- Props and actions:
holding curved scissors,tucking hair behind ear - Character identity (if in blueprint)
- Character/series trigger words should be written according to Anima's Danbooru training set actual token output; do not split character names and series names into multiple tokens.
- When parentheses, colons, and other special characters in trigger words that identify works conflict with Anima's weight syntax, they must be escaped. For example: Danbooru original trigger
37_(reverse:1999)should be written as37\(reverse1999\)in Anima prompts, not as37, reverse:1999,37_(reverse:1999), or37 (reverse1999).
The Tag block is concise, scannable, and does not repeat synonyms. Design logic is not stuffed into tags.
Part B — Natural-language block
One or two coherent English paragraphs, responsible for relationships that tags cannot express:
- Clothing layering, tailoring, inner/outer relationships (
open coat over,hem showing beneath) - Silhouette spatial placement (
behind the head,descending from one shoulder) - Specific distribution of asymmetry
- How the pose affects hem, sleeves, hair, props (blueprint's causal chain)
- Density distribution (
ornament concentrated around ___; ___ left plain) - Light direction, falloff, material response
- Exact punctum location (
the only saturated color is ___ at ___)
The NL block is not a synonym paraphrase of the Tag block. It must supplement relationships and hierarchy.
Template
[comma-separated tags]
[One or two paragraphs: garment hierarchy, silhouette placement, asymmetry, pose causality, density, light, palette hierarchy, punctum location.]
4. Length Budget (for planning, not a passing line)
| Task | Tag + NL total word count |
|---|---|
| Headshot / Bust | 30–50 |
| Character illustration / Fashion | 50–80 |
| Multi-layer display board | 70–100 |
| Narrative scene illustration | 80–120 |
Deletion order when over budget: repeated adjectives → minor accessories → background details → material details. Do not delete: proposition structure, main anchor points, punctum, locked facts, key words in causal chains.
5. Output Modes
Direct Mode (user requests to give directly)
- One line of design direction
- Tag block
- NL block
Standard Mode (default)
- One line of Strategy Line (from blueprint's proposition)
- V1 Faithful: Faithful translation of blueprint, only language conversion, disambiguation, structured formatting; no aesthetic upgrade
- V2 Enhanced: On top of V1, add lighting, material behavior, composition refinement. V2 must not change V1's core facts (hair color, eye color, clothing items, specified pose); if variations are needed, clearly mark as optional variant
- Enhancement Notes: within three lines, explain what V2 added
Deep Mode (when user requests full analysis)
- Proposition, locked info, composition and costume design decisions (citing blueprint)
- Tag + NL
- Brief validation notes
By default, do not output verbose reasoning and aesthetic declarations.
6. Multiple Subjects and Ambiguity
- Multiple characters: group by position (
[LEFT]: ... [RIGHT]: ...), with colors immediately next to nouns within each block. Seereferences/anima-troubleshooting.mddiagnostic item 04. - Unlocked details in blueprint: keep open; do not arbitrarily supplement (do not add earrings, belts, watches to unspecified characters).
7. Feedback Loop
When generated image has issues:
- Image artifacts (hands, color bleeding, background stealing the show, line bleeding, white edge blending) →
references/anima-troubleshooting.md, single-variable minimal fix. - Design issues (too bland, too messy, doesn't look like OC, lacks humanity) → do not fix here. Return to
evaluation-loop→aesthetic-director-core/references/feedback-diagnosis.md. Changing prompt words cannot cure design-layer diseases.
8. Pre-output Checklist
- Is the input a blueprint, not a requirement?
- Is the Tag block concise, specific, without synonym repetition?
- Does the NL block write relationships (hierarchy / position / causality / density / punctum location), not paraphrase tags?
- Are locked facts untouched?
- No vague quality words, no parameters, no default negatives?
- Is length within budget, with minor items deleted?
- Did V2 not change V1's facts?
- Does it comply with the user's requested output format?
References
references/anima-model-profiles.md— Evidence grading, source tracing, model variants, text encoder compatibility, quality word list, parameter reference, experiment logreferences/anima-troubleshooting.md— Artifact diagnostic catalog, minimal fixes, uncertainty handling- Test suite:
tests/test-suite.md