SuperAPI AI Directory
Partner中文AI newsFavorites
Skillsgimi-illustration

gimi-illustration

Design & creation

View repository

Updated 10-15 20:09

About this skill

gimi-illustration · Illustration Skill

Public 3.0 = Product v3.0(multi-style: quirky sketch / warm storybook / product proposal + custom IP)

New window: if there is docs/ first read docs/INDEX.md → docs/roadmap.md; execute based on this file + references/ip/_template.md as authoritative

⚠ Routing (read first · prevent ImageGen from jumping the gun)

When triggered with 自定义 IP / 录入 IP / 上传形象 / 新建 IP(images may be attached):

  1. immediatelyload references/ip/_template.md, follow Step 0；do notfirst treat it as generating images for「illustration / poster」
  2. This turn is prohibited fromcalling any image generation tool (image_gen / GenerateImage / ImageGen, etc.), until the user has answered 「Enroll according to this character design proposal?」
    • same for half-body/full-body: first roundonlyproduce the character design proposal + closing question;zero exceptions
    • After confirmation: only for half-body is the 0.2b full-body design draft allowed; then 0.4 calibration
  3. Prohibitedto understand enrollment as 「generate an IP-themed poster / original illustration poster」
  4. Detailed rules and wording are subject to _template as authoritative; if this section conflicts with it,this section's routing is authoritative(intercept jumping the gun first)

If triggered only by 配图 / illustration etc., and it is not IP enrollment → go to Quick Mode / Step 1–5 below.

Input article content, generate illustrations. Default style 「quirky sketch」(quirky-sketch), switchable to enrolled styles (e.g. 「warm storybook」); IP is gimi / none / custom instance, default 16:9.

Trigger words

Illustration:配图 / 给文章配图 / 帮我配图 / illustration
Custom IP:自定义 IP / 录入 IP / 上传形象 / 新建 IP

Reference files (load as needed)

File When to load
docs/INDEX.md New window document navigation (read first)
docs/roadmap.md New window workbench / when rules need changing
docs/spec.md Product planning (prevent drift); not an execution protocol
references/styles/{$STYLE}.md When generating images / calibrating / style QA
references/styles/_template.md When batch-enrolling new styles (not the illustration hot path)
references/composition-patterns.md When producing strategy (not Step 0）
references/visual-promise.md Step 2 judge visual promise / structure·external parameters·domain objects (SSOT; details are not copied into this file)
references/shot-config.md When producing strategy (not Step 0）
references/prompt-template.md When assembling prompt
references/qa-checklist.md During QA
references/ip/_template.md Custom IP / Step 0 (prioritized over image generation)
references/ip/gimi/ip.md $IP = gimi
references/ip/none.md $IP = none
references/ip/{id}/ip.md $IP is an existing custom id

Image generation capability

Call the current AI softwarebuilt-in image generation tool(e.g. Codex @Image Gen、Cursor GenerateImage , etc.). The Skill handles strategy, prompt assembly, reference image attachments, QA; image generation is executed by platform tools. Interface semantics: text prompt + optional image attachments.

During Step 0 enrollment: follow the 「routing」 at the top of the document; before the allowed conditions are met,must notcall image generation tools.

Platform adaptation priority:

  1. For platforms that support explicit image attachments: pass the reference image as an image reference attachment.
  2. Codex image_gen：For official illustrations of an existing IP, always pass "setting image + current (IP, STYLE) calibration image"; run Step 1.5 first; if either image is missing, official illustrations must not be generated; then image_gen
  3. When image input is not supported and the image cannot be placed into context, pause and explain the reason; for an existing IP, do not downgrade to an official illustration using only a text anchor.

Core variable

Variable Default Description
$STYLE quirky-sketch Default is grotesque hand-drawn; can be switched to a recorded style according to the owner's language (see Step 1)
$RATIO 16:9 Aspect ratio; the user may specify any ratio
$IP gimi gimi / none / Other: if references/ip/{id}/(not _template) exists, load that instance
$REF_MODE Derived by parsing dual / style-single / none; when an existing IP lacks calibration for the current style, it is lazy Gating, not a deliverable reference mode
$TITLE None Handwritten title at the top when requested by the user
$COUNT Automatically inferred Illustration count
$OUTPUT_SIZE None Optional; record only when the user explicitly requests specific pixels

Recommended pixels (not mandatory): 16:9 → 1920×1080 recommended; 3:4 → 1080×1440 recommended.

Workflow

Quick Mode

The user has provided the main text and the variables can be inferred → do not ask further, generate the image directly. Confirm only in the following situations: the user says "do not generate the image yet", variables conflict, a new IP has no ip.mdyet (run Step 0 first), or Step 1.5 judges it as lazy。 Step 0 in progress: turn off Quick Mode(do not skip the confirmation script).
When a recording trigger word appears: Quick Mode image generation does not apply.

Step 0 · Custom IP (as needed)

When the user wants to create/record an IP: load references/ip/_template.md, and strictly follow Step 0 therein to fill the slots (the user only needs to attach images + "record IP"; they do not need to write Step numbers).
Two layers: The protocol follows 0.1–0.5; for the user, use only "user-layer script" (see _template), referring to the current $STYLE 's display_name(the recording default is mostly "grotesque hand-drawn"), and do not expose Step numbers.

Hard gate (before 0.5 is completed) — consistent with the routing at the beginning of the document:

  1. Entry (same for half-body/full-body): In the first round, output only the character design proposal + "Record with this character design proposal?";image generation is strictly prohibited; poster-style final artwork is prohibited
  2. Illustration preemption is prohibited: must not enter Step 1–5; must not write shot-config.md
  3. After image generation, a question must be asked: append a closing question in the same round;image ≠ closure(image generation is allowed only after 0.2 confirmation)
  4. Half-body: after confirmation, 0.2b must be head to toe; when saving, reference-character.png is always full
  5. 0.4 Standard style sample shot (read styles/{$STYLE}.md); before "Does it look right?" is closed, must not proceed to illustrations

examples/{$STYLE}/ If empty and this style is to be used for illustrations → lazily supplement 0.4 (enabled, see _template multi-style details).
Only after completion does it proceed to illustrations; during illustration generation, $IP={id}。

Step 1 · Intent confirmation

Record:$RATIO · $IP · $STYLE · $TITLE · $COUNT · (optional)$OUTPUT_SIZE

Quick Mode inference: vertical/Xiaohongshu → 3:4; otherwise → 16:9.

$STYLE Parsing (master language, user is not required to write an id):

  1. Scan references/styles/*.md(excluding _template.md) of display_name / aliases
  2. explicit hit → lock the corresponding id
  3. soft hit → one-sentence confirmation (e.g., "Do you mean that kind of paper-textured warm color like 'warm-tone picture book'?")
  4. no hit or "change style" → humanized menu: each item display_name + tagline(default can be marked "commonly used")
  5. no style mentioned → quirky-sketch；first image in this sessionappend a light hint (without interrupting):
    "This time we'll use grotesque hand-drawn first. If you want to switch to warm-tone picture book (paper-textured warm color, travel journal feel), just say so."
  6. user describes a style not entered → do not invent on the spot; explain that currently only entered styles are available, ask whether to use the default or enter it later

colloquial first;$STYLE=… advanced only. Refer to the user as display_name, do not mention paths.

Step 1.5 · Current reference parsing (IP and no IP)

Before entering Step 2, writing shot-config.md or generating the formal imagebefore, for the current (IP, STYLE) judge only once:

  1. $IP=none and the current style's "no IP calibration strategy" is required style-single → must follow references/ip/none.md select assets/none/examples/{$STYLE}/ 1 image from, record as $REF_MODE=style-single; missing directory / missing selected image is a release package asset issue, stop and fix, do not ask the user for images, do not downgrade to plain text.
  2. $IP=none and the current style has no required strategy → $REF_MODE=none, keep the existing none behavior; if that style declares optional and has assets, you may follow none.md select 1 image as quality convergence.
  3. already has a full-body character sheet and assets/ip/{$IP}/examples/{$STYLE}/ has at least 1 usable calibration image → $REF_MODE=dual, may proceed to Step 2.
  4. already has a full-body character sheet but the current style's calibration directory is empty / missing files → $REF_MODE=lazy; generate a sample only according to the current style 0.4, after QA ask "Does it look right?";forbiddenwrite shot-config.md, generate the formal image, or attach other styles' calibration images as dual reference. Do not downgrade even if the user urges for the image; should explain "the sample for this style has not been finalized yet, the formal image can only be configured after the sample is confirmed."
  5. character sheet missing or unusable → route to the existing IP entry; must not single downgrade or generate the formal image.

only after the user confirms "looks right" should that sample be saved to the current $STYLE directory; subsequent formal requests re-parse as dual。formal image configuration for an existing IP is not allowed to be single, must be dual. IP files do not maintain a global ref_mode; the current style's directory facts are the sole basis.

Step 2 · Digest content + produce strategy

only $REF_MODE has been judged by Step 1.5 as dual / style-single / none can enter. Strategy:composition-patterns.md · Visual promise:must load visual-promise.md(SSOT) · Save to disk:shot-config.md。

  1. load composition + shot-config + visual-promise
  2. according to composition Step 2 decision flow(two axes → creative generation method eight steps → composition library)
  3. each shot fill the visual promise full set (visual-promise.md）+ density / reading order(composition "reading flow grammar";scenic mandatory)
  4. 外参参考: Per visual-promise §B (main anchor budget + default feature lock without attached images); on failure/timeout silently degrade 意象(without asking the user)
  5. Quick Mode: by default save to disk and proceed to Step 3; when the user says "show me the strategy first", display the summary first

Step 3 · Assemble Prompt + Generate Image

  1. Load styles/{$STYLE}.md、prompt-template.md, corresponding ip/*.md
  2. only consume what was already parsed in Step 1.5 $REF_MODE：
    • dual → character setting image + examples/{$STYLE}/ 1 calibration image from (both must exist to assemble)
    • style-single → $IP=none, must pass assets/none/examples/{$STYLE}/ according to none.md the selected 1 image; the prompt explicitly states to match only the drawing style, not to copy the composition, title, object arrangement, or business content
    • none → still follow the none rule (no character setting image, and no required style assets)
    • Reference images did not enter the context; must not claim that an IP process was executed
  3. Per visual-promise.md / prompt-template: append the promise lock;外参参考 by default do not attachlandmark image (features go into SUBJECT); attach only when "external reference attached = yes";scenic append the reading flow implication (composition SSOT)
  4. Fill in the variable slots, call the image generation tool (IP_DESC before STYLE_DNA , see prompt-template.md）

Step 4 · QA Quality Check

Load qa-checklist.md(promise → visual-promise; flow → composition "reading flow grammar").$OUTPUT_SIZE When there is a value → proportional scaling + white background padding, cropping prohibited. If it fails → retry, at most 2 times.

Step 5 · Delivery

outputs/{YYYYMMDD}-{slug}/
  shot-config.md      ← 配图策略 + 插入位置（与图片同时交付）
  01-{slug}.png
  02-{slug}.png       ← 多张时继续编号

Version

Public / Product 3.0 · Three styles selectable + custom IP · visual-promise SSOT