Halloween Elf Witch — Open Skill
Halloween Elf Witch
Halloween Elf Witch is a public Open Skill on Digen. Open it to see what this workflow creates.
TryOn
0 0 0
https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/4284865/2026-10-22/pvpzthn7.jpg
# Halloween Elf Witch
Text-to-image Skill that turns a short text description into one Halloween dark-fantasy
illustration of an elf witch. There is no structured project state: everything the Skill
needs is in the user's message plus the two reference files below.
## When to use
- The user describes an elf witch / witch-elf / "bruxa elfa" scene and wants an image.
- The user asks for a Halloween witch artwork with pumpkins, a haunted castle, a full moon,
a pointed hat, elf ears, or a dark-fantasy portrait poster.
- The user asks to redo, restyle or vary a previously generated elf-witch image
(different pose, outfit, closer framing, more props) while keeping the same character.
## Steps
1. **Read the visual rules.** Call `read_skill_file` for
`references/style-and-prompt.md`. Never write the image prompt from memory — the
style, subject, layout, must-keep and must-avoid blocks live in that file.
2. **Read the assembly rules.** Call `read_skill_file` for `references/step-details.md`
to get the input-parsing checklist, prompt assembly order, branching rules,
default parameters and the retry conditions.
3. **Build one English prompt.** Follow the template in `references/style-and-prompt.md`:
start from the fixed Halloween dark-fantasy style and layout blocks, then insert the
user's subject description (pose, outfit, expression, extra props). Keep the
must-keep list and append the must-avoid list. Keep it roughly 60–140 words,
single paragraph, English only.
4. **Generate the image.** Call:
`generate_image(prompt=<the prompt from step 3>, model='t2i', orientation='portrait', resolution='1080P', name='Halloween Elf Witch', entity_type='characters')`
- `model='t2i'` is the fixed text-to-image channel for this Skill. Do not pass a
physical model name and do not switch channels unless the user supplies a
reference image to edit — only then use `i2i.hd.lite` with `input_urls=[that image]`.
- `orientation='portrait'` is the default because the artwork is a vertical poster;
switch to `landscape` or `square` only if the user explicitly asks for that shape.
- `resolution='1080P'` unless the user asks for higher or lower detail.
- Use a short `name` that reflects the request (e.g. `Halloween Elf Witch — Moonlit Gate`).
5. **Reply briefly.** One or two sentences in natural language describing what was made
(character, scene, mood). Do not paste long URLs and do not describe the prompt verbatim.
## Notes
- If the user only gives a subject ("an elf witch"), still keep the full Halloween scene
from the reference file; do not fall back to a plain studio portrait.
- Keep the character identity stable across variations: elf ears, pointed witch hat,
long wavy blonde hair, black lace dress and cloak are must-keep unless the user
explicitly changes them.
- If `generate_image` returns `success=false`, say so honestly, do not invent an image
or a link. Retry at most once after applying the fixes in `references/step-details.md`.
- Refuse requests for nudity, sexualized minors, gore or real-person likeness; do not
call the generation tool for those.
- Do not add invented facts about the original app (prices, artists, features) — this
Skill only reproduces its look.