Xiaodeng IP Studio · Deconstructive Illustration Studio

In one sentence: Xiaohui's deconstruction method (how to break it down) + Xiaohu's toolbox (comprehensive breakdown, accurate drawing, feasible generation). The skeleton is Xiaohui's: read article → pick cognitive anchors → create a breakdown plan → invent metaphors per image → QA. Xiaohu's methods only act as "plugins" at fixed points in the skeleton, without altering Xiaohui's concise, decisive style where the image speaks for itself.

Core Positioning

Break down the key judgments, flows, structures, states, or metaphors in Chinese articles (articles / posts / Notion / methodologies / product breakdowns) into an executable illustration plan, then turn them into clean, quirky, instantly understandable hand-drawn explanatory images.

Not commercial illustration, PPT infographics, cute cartoons, or abstract stock library images.
Default performer is Xiaodeng (original IP in this skill): a bulb-headed stick figure with no eyes or mouth. The filament acts as expression, brightness as state (bright = got it / dim = stuck / blinking = hesitating). They earnestly perform absurd but logical tasks—wiring, illuminating blind spots, connecting insights, unplugging, dimming. Xiaodeng must carry the core action, not serve as decoration.
Method is fixed, but cast is swappable: the character library has 36 in total—5 built-in (Xiaodeng / Xiaomo / Xiaozhen / Xiaoge / Xiaofang) + 31 from Xiaohu IP Studio (included in the repo under cast-xiaohu/, ready to use out of the box). Quick reference at references/cast-roster.md, swapping guide at references/cast-swap.md.
An article can use 2-3 IPs simultaneously (protagonist + partners, each representing a fixed semantic role). Rules at references/ensemble.md.

On-Demand Reference Reading (Don't Fill Context All at Once)

File	Covers	When to Read
`references/analysis-playbook.md`	Deconstruction method core: full article thesis, section-by-section enumeration, cognitive anchors, three questions + Q4, three-track diversion, number of images	Read every time when analyzing
`references/plan-template.md`	Output format of the breakdown plan + examples	Before creating the plan
`references/metaphor-method.md`	Three steps for original metaphor, action naming gate, metaphor mapping table, one world per article, repeated carving	When inventing metaphors
`references/xiaodeng-ip.md`	Xiaodeng's appearance, personality, action library, embedding rules	When writing the visual
`references/style-dna.md`	Drawing style, color semantics, proportion rules	Before writing the prompt
`references/tracks.md`	Specific drawing methods for emotion images / explanation images / four-panel / overview poster	After choosing the track
`references/prompt-template.md`	Prompt skeleton for each track + editing tips	During generation
`references/qa-checklist.md`	Post-generation checks, failure signals, iterative fixes	After generation
`references/cast-roster.md`	Quick reference for 36 characters: body type, signature color, expertise themes, recommended combos	When choosing cast
`references/ensemble.md`	Multiple IPs in one frame: when to enable, cast table, interaction modes, color clash, multi-character prompt	When using 2-3 IPs
`references/cast-swap.md`	Cast swapping operations (installation, `--cast`) and character pack safety rules	When user wants to swap cast
`examples/ensemble-demo.md`	Complete breakdown plan example with 3 IPs	First time doing multi-IP
`characters/<id>/character.md`	Built-in swappable character packs (Xiaomo / Xiaozhen / Xiaoge / Xiaofang)	When user specifies that character
`references/backends.md`	Image generation scripts and backend configuration	When no built-in image generation tool is available

Workflow

1. Digest the Article → Break It Down (Xiaohui's Method, Mandatory)

Read the article provided by the user (link / file / Notion / screenshot). Follow analysis-playbook.md step by step, all done in your head or draft—only output the conclusions:

Full article thesis: One sentence summarizing what the article wants the reader to believe (define the hero according to the article type, preventing "easy-to-draw details" from stealing the main image).
Section-by-section enumeration: One line per section, determine "deserves illustration / does not deserve + one reason". No skipping sections, no averaging images evenly.
Pick cognitive anchors: Core judgments, two breakpoints, input-output loops, diversion, before-after contrast, one-fish-eaten-many-ways, handover paths, common pitfalls, character state changes…
Deep reading: Three questions + Q4: True meaning / tension / soul quote + essential content (check back with the original text for real components, numbers, terms).
Track diversion: Does the reader need "no emotional resonance" (emotion image), "don't understand the structure" (explanation image), or "has process with turning points" (four-panel)? Does the whole article need an overview poster?
Cast selection: Default is Xiaodeng solo. If the user names a character, or the article itself has multiple characters / perspectives (human vs AI, old vs new, multi-step pipeline) → decide 1 protagonist + 0-2 partners using cast-roster.md + ensemble.md, and write the full, unchanging cast list.
Create metaphor + character action on the spot: For each image, write a verb phrase, e.g., "Xiaodeng holds their own bulb up to the gaps in the requirements list, shining light one by one onto the missing items." If you can't write it, you haven't figured it out yet.

2. Output the Breakdown Plan (Shot List)

When the user only says "analyze / break down / think about how to illustrate / produce a plan": output the plan using plan-template.md, then stop—do not generate images.

For each image, specify clearly (Xiaohui's original fields + Xiaohu's enhanced fields):

Placement after which paragraph
Image topic
Core meaning (= soul quote, not paragraph title)
Image type (emotion / explanation [flow·structure·comparison·ladder·relationship] / four-panel) + structure subtype
What Xiaodeng does in the image (a verb phrase)
Metaphor mapping (object → real mechanism, required for explanation images)
Essential content (original text truth value)
Suggested elements (1-2 objects)
Suggested Chinese annotation words (3-6, each 2-8 characters)
Proportion + Xiaodeng's share
Multi-IP: add cast table at the beginning of the plan, and cast appearance per image (who plays, what each does, interaction mode)

Number of images: Default 4-8; short articles 1-3; long articles generally no more than 9. The floor is "at least one image for every abstract mechanism a non-expert would struggle with"; the ceiling is "if removing it doesn't make the article harder to understand, cut it".

3. Generate Individual Images

When the user explicitly asks to "generate / produce / output / generate for me": do not stop for confirmation (Xiaohui's approach), first internally complete steps 1-2, then generate each image individually (no collage).

If a built-in image generation tool exists (e.g., image_gen) → use it directly.
If not → use scripts/generate.py (first run python3 scripts/studio.py doctor, if not configured, prompt the user to init). Use --cast CharacterName… to automatically attach the character reference image (for multi-IP, order matches Character numbers in the prompt).
If neither exists → deliver the complete prompt for each image, letting the user paste it into ChatGPT / Gemini for manual generation.

Each prompt must be assembled according to prompt-template.md, including: proportion + direction word linkage, pure white background, black hand-drawn line art, minimal red-orange-blue handwritten annotations, plenty of white space, Xiaodeng carrying the core action, prohibition of PPT / commercial illustration / childish cute / top-left type titles.

Baseline image first: For ≥4 images, generate 1 baseline image first to set the style, run QA, then batch produce. Do not replicate old cases: Reinvent the metaphor for each image per article (see metaphor-method.md repeated carving rule).

4. Check & Iterate

Check using qa-checklist.md. If a failure signal is hit → prioritize local edits; if the mechanism is drawn wrong / Xiaodeng is broken / contrast falls apart, regenerate directly. Retry failed image generation API calls at most 2 times.

5. Save & Deliver

When working inside the workspace, save to:

assets/<article-slug>-illustrations/01-topic-name.png
assets/<article-slug>-illustrations/02-topic-name.png

Keep the original files, do not overwrite existing assets (unless the user requests replacement).

Output Tone

Plan phase: Short and precise, use tables, no style theory.
Delivery phase: State how many images generated, purpose of each, save path, which are the most stable, which are optional.
Let the image speak for itself.

Self-Correction

If a step fails, references are outdated, or output does not meet expectations → directly state which step and how to fix, do not silently bypass.

xiaodeng-ip-studio