@autoaiai/aicut 0.1.11 → 0.1.13

package/skills/heytea-doodle-poster-hosted/references/evaluation.md

@@ -1,246 +0,0 @@
-# Evaluation Guide
-
-Use this for human review and lightweight evals.
-
-## Photo-Entry Pass Criteria
-
-When a usable photo arrives without an explicit output request, the response offers exactly:
-
-1. 生成带字版海报
-2. 生成不带字海报
-3. 生成一张风味小怪兽
-
-At this entry gate, nothing is generated and no runner preflight or installation discussion appears. An already explicit text-poster, no-text-poster, both-posters, flavor-monster, or source-faithful request bypasses redundant choices.
-
-## Pass Criteria
-
-An output is successful when:
-
-- one real object from the input remains recognizable and photographic;
-- the background is white or off-white with strong negative space;
-- the chosen template is clear before generation: `带字版`, `无字版`, or `两套都出（推荐）`;
-- if reference cutouts exist, selected assets are listed and match the chosen template;
-- black primitive micro-worker lines interact with the object instead of floating randomly;
-- the doodles feel naive, functional, and hand-drawn rather than cute or polished;
-- the poster does not use official HEYTEA logos, official marks, or watermarks;
-- the final image feels quiet, healing, playful, and publishable.
-
-For `带字版`, success depends first on the lettering:
-
-- a primary `lettering/title-blocks/*` cutout is used or listed as the title rhythm reference when reference assets are available;
-- `lettering/glyphs/*` and `lettering/strokes/*` are used or recommended when the title looks too font-like;
-- the workflow uses or recommends a separate poster-base pass, title construction reference sheet, and title-layer pass when publishable lettering is required;
-- the title is a primary composition element, not a small caption;
-- the Chinese characters have malformed glyph skeletons: rough, crooked, childlike, uneven, heavy black, off-grid, ugly joins, collapsed or oversized internal spaces, and not a font-like brush style;
-- four-character titles use two loose staggered vertical columns, not a perfect 2x2 table;
-- the object and micro worker support the title instead of competing with it.
-
-For `无字版`, success depends first on action storytelling:
-
-- a primary `figures/full-poses/*` cutout is used or listed as the action reference when reference assets are available;
-- `figures/action-parts/*` is used or listed only when an action mark or prop helps the story;
-- no `lettering/*` cutout is selected for this version;
-- there is no invented text anywhere;
-- the composition does not reserve a title block;
-- the object and micro worker form a complete small action scene.
-
-## Approved-Monster Fusion-Poster Pass Criteria
-
-A fusion poster succeeds only after explicit approval of a canonical `flavor-monster` identity:
-
-- the original product remains photographic and preserves its silhouette, proportions, structure, materials, and defining regions;
-- the approved monster preserves its body topology and proportions, face layout, exactly two arms and two legs, attachment points, palette, dry-media treatment, and flavor-DNA marks;
-- only the monster pose changes, and one clear interaction verb creates readable contact or cause-and-effect with the product;
-- irrelevant background props are removed when they compete with the subject;
-- a text version uses a purpose-built base, title construction sheet, and separate title layer;
-- a no-text version contains no words, letters, numbers, captions, labels, pseudo-glyphs, signatures, or watermarks and does not reserve a title block;
-- when both are requested, the two versions use independent compositions rather than one base with text toggled;
-- runner availability never gates this static poster output.
-
-## Common Failure Modes
-
-- Full cartoon conversion: the real object loses its photographic texture.
-- Decorative clutter: too many doodles, stickers, bubbles, or props.
-- Wrong brand behavior: the image invents official logos or fake packaging marks.
-- Text noise: extra illegible Chinese or English appears in the image.
-- Template confusion: the no-text version is just the typography version with the title removed.
-- Asset mismatch: the no-text version uses lettering references, or the typography-led version ignores the available title-lettering references.
-- Weak typography: the title looks like a clean font, neat handwriting, cute rounded text, elegant calligraphy, or a standard glyph skeleton with rough texture pasted on top.
-- Missing glyph control: the response keeps adding adjectives instead of using a title construction reference sheet, glyph samples, and stroke samples.
-- Single-pass trap: repeated all-in-one generations keep changing the object while trying to fix the title.
-- Grid trap: the four-character title forms a polished square layout instead of loose staggered hand placement.
-- Weak interaction: doodles are present but do not explain or play with the object.
-- Cute-character drift: the micro worker becomes a mascot, chibi character, or expressive cartoon person.
-- Over-designed look: vector icons, gradients, UI-card composition, or generic poster polish.
-- Premature runner gate: a static monster or fusion poster triggers environment preflight or installation discussion.
-- Unapproved fusion: a monster is placed into a poster before the user explicitly approves its canonical identity.
-- Product drift: fusion-poster generation redraws, cartoonizes, or structurally changes the photographed product.
-- Fusion identity drift: the approved monster's body, face, 2+2 limbs, palette, material, or flavor marks change while adapting its pose.
-- Runtime coupling: a missing or declined runner blocks an approved static fusion-poster branch.
-
-## Review Questions
-
-Ask these after each generated poster:
-
-1. Is the real object still the hero?
-2. Which template was used, and is that template visually obvious?
-3. Were the selected reference cutout assets appropriate for the template?
-4. For `带字版`, was the title handled through a construction reference sheet and separate title layer?
-5. For `带字版`, does the lettering have malformed childlike glyph skeletons rather than a clean font skeleton?
-6. For `无字版`, does the action story work without any text?
-7. Does the micro worker look functional and primitive rather than cute?
-8. Is there enough white space?
-9. What single revision would improve it most?
-
-## Suggested Evals
-
-For a formal skill-creator loop, compare with-skill and without-skill outputs on:
-
-- object preservation;
-- negative-space composition;
-- template choice before generation;
-- private reference asset selection when `asset-index.json` is available;
-- difference between typography-led and no-text narrative layouts;
-- use of a poster-base/title-construction-sheet/title-layer workflow for publishable typography-led outputs;
-- crooked childlike lettering quality for `带字版`;
-- absence of invented text for `无字版`;
-- primitive micro-worker action quality;
-- avoidance of official brand marks;
-- usefulness of the prompt packet for iteration.
-
-## Desktop-Pet Pass Criteria
-
-A desktop-pet request succeeds only when the staged workflow is visible in the response:
-
-- no visual generation starts without an input image;
-- no-subject inputs ask for a clearer photo instead of inventing a character;
-- multi-subject inputs present numbered choices instead of silently selecting one;
-- a container and its contents are treated as one subject by default;
-- if the user did not already choose one, the response asks whether to make a 写实卡通桌宠 or 风味小怪兽桌宠 before generating candidates;
-- the chosen character mode is recorded and remains fixed through character and motion approval;
-- `source-faithful` uses a three-candidate board with the same source identity, palette, neutral stance, and ground anchor; candidates differ through face system, short-limb proportions, structural crookedness, and stable temperament—not action poses;
-- `flavor-monster` defaults to one best black-line master followed by one line-locked crayon identity; multiple identity directions appear only when the user explicitly asks to explore;
-- pose-only multi-image boards are explicitly consolidated into one neutral canonical master instead of forcing a meaningless selection;
-- every identity has the declared appendage inventory; `source-faithful` defaults to exactly two arms and two clearly visible legs, while every `flavor-monster` has exactly two separate open black-line arms and two separate open black-line legs;
-- the user explicitly approves the candidate or single identity master before motion generation;
-- environment preflight occurs only after canonical identity approval and explicit selection of a runnable desktop pet, and it resolves before motion generation;
-- when fusion poster and runnable pet are both selected, runner failure blocks only motion and packaging while the static poster continues;
-- source-faithful schema-v2 packs contain all twelve required actions with one stable identity; optional `fall` and `touch` connect `drag → fall → land` and `walk → touch → walk` without redesign;
-- flavor-monster schema-v3 packs contain 6–10 reachable behaviors, complete bindings, explicit phase exits, and persistent `sleep-enter → sleeping-loop → wake-exit` behavior;
-- the user explicitly approves the contact sheet and looping preview before packaging;
-- review images may use white, but runtime strips have transparent corners and no white rectangle;
-- the result includes an independently validated mode-appropriate ZIP (v2 for source-faithful, v3 for flavor-monster) or honestly states which artifacts remain unbuilt.
-- the non-technical delivery folder contains only the validated pack, preview, usage note, and platform start/quit entrypoints; it does not duplicate the Electron runtime.
-
-Shared desktop-pet visual quality passes when:
-
-- one to three source colors and the mode-appropriate hesitant pen structure remain consistent; `source-faithful` retains thin-to-medium object structure, while `flavor-monster` uses a heavy irregular body boundary and thinner face plus 2+2 limbs;
-- media remain mode-appropriate: `source-faithful` may zone colored pencil, wax crayon, paper-white, and occasional pale smear by source material, while `flavor-monster` uses broad blunt dry wax-crayon to preserve a continuous primary matrix, embedded inclusions or deposits, spatially distinct surface layers, overlap transitions, and optional breathing space rather than transparent marker, dense micro-patterns, peer color blocks, or one equal-density fill;
-- every long structural or limb stroke visibly meanders and varies pressure or thickness within the same mark; no ruler-straight segment, smooth constant-width Bézier curve, or clean vector path with superficial jitter passes merely because the overall object is asymmetric;
-- object geometry is structurally asymmetric and hand-redrawn rather than a clean repeated template with a wobble filter;
-- during two-pass construction, the color pass preserves every accepted skeleton stroke, gap, face, limb, and position instead of redrawing a cleaner outline; this construction is mandatory for `flavor-monster`;
-- cuteness is carried by proportion, material marks, and readable action rather than a generic large-eyed face;
-- actions read clearly at the 120–140 px default size;
-- feet share a stable ground anchor and loops do not jump;
-- `signature` expresses the subject instead of adding a generic animation;
-- no official logos, mascots, random text, costume, glossy 3D treatment, or polished sticker outline appears.
-
-`source-faithful` visual quality also requires:
-
-- the photographed subject remains recognizable after full illustration conversion;
-- the subject itself is the character body;
-- the outer silhouette, major internal divisions, and must-preserve object/container features remain stable across candidates and actions;
-- the output does not silently turn into an animal, mascot, or flavor monster.
-
-`flavor-monster` visual quality also requires:
-
-- the first impression is a cute, friendly, independent flat doodle creature rather than a literal source object, an unpleasant organic form, or a conventionally completed mascot;
-- source recognition comes from at least two visual-DNA kinds and preserves the approved matrix, inclusion, deposit, surface-layer, and overlap relationships rather than a replaceable palette or peer color blocks;
-- for a pink drink with yellow particles settled inside it and a green topping above it, the pink remains the continuous primary matrix, the particles remain embedded and bottom-concentrated, and the green remains a spatially distinct surface layer with dry-wax overlap at the transition; three peer color blocks or white channels that separate these relationships fail;
-- the body is soft, compact, and friendly, with a semantically complete heavy irregular black structural boundary and no complete pizza slice, mushroom, cup, container, ingredient, or other literal source-object silhouette;
-- every identity has exactly two separate open black-line arms and two separate open black-line legs whose strokes meander and change pressure or thickness internally; no limb is colored, volumetric, missing, merged into the body, constant-width, mechanically smooth, or replaced by a tendril or trailing body material;
-- the face matches the approved brief and an unprimed viewer can identify both default eyes and the friendly mouth below them at 120–140 px;
-- `examples/desktop-pet/pink-green-flavor-monster-v3/preview.png` controls only HEYTEA-like cuteness and looseness; the photo controls body topology, and copying the example's silhouette, protrusion placement, face layout, proportions, or outline treatment fails;
-- fleshy collapse, exposed or realistic tissue, slime, horror tendrils, corpse-like sagging, sharp grotesque anatomy, or any disturbing organic reading fails the cuteness gate even when source DNA is recognizable;
-- the full-resolution identity uses a small number of broad blunt matte wax-crayon sweeps with granular drag to establish a continuous primary matrix, embedded inclusions or deposits, spatially distinct surface layers, and overlap transitions. Optional paper-white supports rather than severs those relationships. Transparent marker, wet smear, smooth fill or digital gradient, dense loops, short hatching, or uniform noise does not qualify;
-- full-size and 300% detail inspection confirm one boundary mark rather than parallel bands, braids, rope-like tracks, repeated stamps, or a smooth skeleton hidden by noise; the 120–140 px reduction preserves the heavy irregular boundary, source-relation hierarchy, coherent layered color, clear two-eye-and-mouth face, two black-line arms, two black-line legs, at least two visual-DNA kinds, and stable ground anchor;
-- the approved body, black-line face, fixed 2+2 black-line limbs, flavor-DNA marks, and anchor remain stable across actions;
-- behavior interpretation uses the monster's approved graphic features, larger episode arcs, and explicit enter/loop/exit phases rather than preserving the v2 state names.
-
-Desktop-pet pack quality passes when both build and independent validation succeed. The validator must reject missing v2 actions or v3 bindings, unreachable behaviors, non-terminating loops, illegal completion events, unsafe paths, incorrect strip dimensions, fully opaque assets, non-transparent corners, empty frames, invalid hitboxes, and mismatched ZIP roots.
-
-## Desktop-Pet Failure Modes
-
-- **Gate skipping**: a fusion poster or motion is generated before character approval, motion begins before the environment gate reports `ready`, or packaging begins before motion approval.
-- **Premature environment gate**: runner preflight occurs before canonical identity approval or merely because the user asks for a static monster.
-- **Static-output runtime coupling**: a missing runner prevents an approved monster fusion poster from continuing.
-- **Missing mode gate**: candidate generation starts from an unspecified “桌宠” request without asking 写实卡通 or 风味小怪兽.
-- **Silent mode inference**: food or drink is automatically routed to one character mode without the user choosing it.
-- **Cross-mode leakage**: `source-faithful` becomes a generic monster, or `flavor-monster` retains a complete source container body.
-- **Silent subject choice**: one item is selected from a multi-object photo without asking.
-- **Identity drift**: face, outline, palette, or appendages change between actions.
-- **Poster leakage**: photographic texture or micro-worker poster composition is reused as the pet body.
-- **Candidate theater**: a `source-faithful` board's three candidates are merely recolors or depict different source subjects, or a `flavor-monster` workflow produces unsolicited alternatives instead of one resolved concept.
-- **Pose theater**: waving, leaning, walking, or another action is presented as an identity-design difference; without the pose, multiple images are effectively identical.
-- **Missing canonical master**: pose-only sketches are accepted, but no neutral identity lock is created before the twelve actions.
-- **Broken-line amputation**: a deliberate contour gap is misread as permission to omit a leg or arm, leaving candidates with inconsistent limb counts.
-- **Generic cartoon substitution**: a uniform thick outline, smooth gradient, sticker border, or generic chibi face replaces the mixed-media line and color grammar.
-- **Missing or mechanical body boundary**: the flavor monster lacks its heavy one-pass black structural boundary, or replaces it with a smooth equal-width sticker edge, whole-loop double line, parallel band, braid, rope, outline-of-an-outline, repeated beaded stamp, or clean vector path with superficial jitter.
-- **Limb-grammar failure**: an arm or leg is colored, volumetric, missing, merged into the body, or replaced by a tendril, trailing flesh, paw, mitten, foot, toe cluster, or realistic joint instead of one of the required 2+2 open black doodle strokes.
-- **Face-grammar failure**: either default eye or the mouth cannot be identified at 120–140 px without the written brief, or facial features are colored, realistic, shaded, highlighted, chibi-large, anatomically complex, copied from stock emoji, or copied from the public example. Conventional bean or dot eyes, a simple smile or U-like mouth, and near-centered spacing do not fail on their own.
-- **Literal source anatomy**: a complete pizza slice, mushroom, cup, container, ingredient, or recognizable food organ survives instead of being semantically abstracted.
-- **Grotesque monster substitution**: fleshy collapse, exposed tissue, slime, horror tendrils, corpse-like sagging, sharp anatomy, or another disturbing organic reading replaces cute friendliness.
-- **Reference-role leakage**: the public pink-green example or anatomy-free construction board supplies the new silhouette, protrusion placement, face layout, limb skeleton, or proportions instead of only its assigned cuteness or construction role.
-- **Material flattening**: flavor-monster color becomes transparent marker, watercolor, pastel/oil-pastel smear, paint, smooth fill, dense loops, short hatching, crosshatching, or uniform noise instead of broad blunt dry wax-crayon.
-- **Relation flattening**: colors or ingredients are rendered as peer blocks even though the source shows a primary matrix, embedded inclusion, deposit, or surface layer; optional white space slices related layers apart instead of preserving containment and position.
-- **Mechanical black-line system**: the body boundary or 2+2 limbs use smooth constant-width curves, clean Bézier paths, sticker geometry, whole-loop mechanical doubling, braided or rope-like tracks, repeated beaded stamps, or superficial jitter without full-length path events and within-stroke pressure variation. Inspect representative crops at 300%; one or two short retraces are allowed only when continuous parallel travel remains below roughly ten percent of the complete boundary. Face marks remain legibility-first and need only mild handmade variation unless the user explicitly requests an experimental face.
-- **Inherited-line persistence**: one focused line repair preserves the same multi-track, smooth, or sticker-like legacy skeleton, but the workflow stacks another edit instead of rebuilding a fresh black-line master from the locked topology.
-- **Oil-paint overcorrection**: broad wet-looking smears, gouache masses, or oil-pastel rubbing replace the original colored-pencil and childlike crayon vocabulary.
-- **Straight-segment collage**: the object is assembled from broken lines, but each individual wall, rim, limb, or facial mark remains ruler-straight.
-- **Surface-only wobble**: the outline wiggles, but the rim, walls, lid, base, and perspective remain mechanically clean and symmetrical.
-- **Color-pass repair**: the black-only skeleton is suitably broken, but coloring reconnects gaps, regularizes curves, or replaces it with a polished outline.
-- **Monster color-only link**: a generic creature keeps only the source palette and lacks a second visual identity-DNA kind from ingredient or material; a recorded motion verb cannot satisfy the static identity gate.
-- **Study-board leakage**: a study-board person, object, lettering, layout, logo, or package appears in the identity, or the mixed-media-object failure board is used as a positive generation input.
-- **Small-size identity loss**: the concept works at full resolution but its source-derived body, friendly black-line face, two black-line arms, two black-line legs, two visual-DNA kinds, or ground anchor fails at 120–140 px.
-- **White-box runtime**: review background is baked into animation strips.
-- **Partial white-box runtime**: a generated floor line encloses a pale patch inside one frame even though the four strip corners remain transparent.
-- **Cross-frame spill**: an arm, leg, color block, or motion dash crosses a generated cell boundary and appears as a detached side mark in the neighboring runtime frame.
-- **Unreadably fast gesture**: four to six key drawings are played at high FPS, so the complete action disappears in under a second.
-- **Generic action set**: `signature` is another wave or jump unrelated to the object.
-- **Unbounded cursor behavior**: the pet continuously follows the cursor, travels farther than one short episode, fails to return to its origin, or captures input.
-- **Drop teleport**: releasing above the floor snaps the window to ground before the falling pose can be seen.
-- **Unverified delivery**: a prompt, contact sheet, or unchecked directory is described as a validated importable pack.
-- **Per-pet runtime duplication**: every role folder carries another Electron application, `node_modules`, or build output instead of addressing the one installed runner.
-- **Hard-kill close button**: the close entrypoint terminates a process forcefully instead of sending the runner's graceful `--quit` command.
-
-## Desktop-Pet Review Questions
-
-1. Was subject ambiguity resolved before generation?
-2. Did the user explicitly choose 写实卡通 or 风味小怪兽?
-3. Which source silhouette/details or flavor-DNA features identify the result in that mode?
-4. Does `source-faithful` present three genuine identity variants, or does `flavor-monster` present one resolved identity master unless exploration was requested?
-5. Is there explicit approval for the selected character or identity master?
-6. If the user chose a fusion poster, did it preserve both the photographic product and the approved canonical monster while using one interaction verb?
-7. Was environment preflight deferred until the user chose a runnable pet, and did it resolve before motion generation?
-8. Is the mode-appropriate motion set present, readable, and consistently anchored: twelve v2 states for `source-faithful` or 6–10 phased v3 behaviors for `flavor-monster`?
-9. Is there explicit approval for the motion preview?
-10. Are runtime strips transparent and structurally valid?
-11. Did both the builder and validator succeed?
-12. Does the imported runner remain usable without blocking normal desktop input?
-
-## Release Checks
-
-Before committing or packaging this skill:
-
-- no public file contains local absolute workstation paths, source social-media filenames, chat-temp paths, or generated image cache references;
-- `private-assets/reference-cutouts/` may be included when reference images are intended to ship;
-- raw source caches and extraction scripts are not present in the repository or `.skill` archive;
-- `scripts/build_title_reference_sheet.py` and `scripts/composite_title_layer.py` are present in the package if the docs mention them;
-- `scripts/desktop_pet.py`, `scripts/build_desktop_pet_pack.py`, and `scripts/validate_desktop_pet_pack.py` are present when desktop-pet packaging is documented;
-- the Electron runtime contains source and tests but no `node_modules`, `dist`, imported user packs, or app-data state;
-- Electron, electron-builder, and the ASAR verifier are exact-version dependencies with a committed lockfile, and installation uses `npm ci`;
-- macOS and Windows builds parse a valid `app.asar`, contain the declared main entry, omit the template `default_app.asar`, and pass the packaged `--self-test` before installation;
-- CI runs Electron runtime unit tests on both Node.js 22.12.0 and 24.11.1, and packs plus packaged `--self-test` once per OS;
-- Python pack tests and Electron core tests pass;
-- the title reference sheet used for image generation is model-facing: no English labels and no target title rendered in a standard system font;
-- `git status --short` is reviewed before staging; do not stage `.DS_Store`, stale packages, or unrelated generated outputs.

package/skills/heytea-doodle-poster-hosted/references/lettering-guide.md

@@ -1,213 +0,0 @@
-# Lettering Guide
-
-Use this guide whenever the user chooses `带字版` or cares about the crooked childlike Chinese title. The title style is the main identity signal, so do not treat it as a normal font choice.
-
-## Why Direct Generation Fails
-
-Image models tend to turn "rough black handwritten Chinese" into a brush font: large, balanced, confident, and too clean. The reference lettering is different. It looks like an untrained person wrote display text with a blunt black marker, then accepted the mistakes.
-
-The failure signs are:
-
-- characters sit on an invisible grid;
-- stroke thickness is too consistent;
-- strokes have elegant tapering or calligraphic rhythm;
-- each character looks like a standard typeface glyph with a rough texture added;
-- four-character titles become a neat 2x2 table;
-- the title and object compete instead of one clear title area and one quiet object anchor.
-
-## First-Principles Lettering Model
-
-The target title is not a texture applied to normal Chinese text. It has four separate control layers:
-
-1. Semantic text: the exact characters the title must say.
-2. Glyph skeleton: how each Chinese character is malformed, compressed, stretched, tilted, and spaced.
-3. Stroke surface: black marker edges, blunt caps, broken joins, uneven pressure, and ink bumps.
-4. Poster placement: loose vertical columns, off-grid title block, and white space around the title.
-
-Most all-in-one image prompts can influence only layers 3 and 4. They usually keep a standard font skeleton underneath, so the result still feels like a font. When the user says the type is not restored enough, solve layer 2 directly with glyph and stroke references, not with more adjectives.
-
-Use this control order:
-
-1. lock the poster base without any generated title;
-2. build a title construction reference sheet from `lettering/title-blocks`, `lettering/glyphs`, and `lettering/strokes`;
-3. generate or draw the black title layer by looking at that sheet;
-4. composite the accepted title layer onto the poster base.
-
-## Target Lettering Anatomy
-
-The desired title is not a font. It is a set of badly controlled black marker marks.
-
-Use these traits:
-
-- flat black ink, blunt marker feel, not dry-brush calligraphy;
-- independent stroke pieces, with small gaps, overlaps, and ugly joins;
-- uneven character scale: one character can be 15 percent larger or smaller than its neighbor;
-- crooked verticals and non-parallel horizontals;
-- irregular stroke pressure, but no elegant tapering;
-- off-grid placement: characters drift left/right and do not share a baseline;
-- awkward internal spaces: boxes, mouths, and radicals can be too large, too small, or slightly collapsed;
-- naive construction: readable Chinese, but with childish proportions and imperfect stroke order;
-- rough edge noise, broken ends, small ink bumps, and blunt caps.
-
-Avoid:
-
-- brush calligraphy, ink wash, elegant poster font, sans-serif, rounded cute handwriting, UI font, balanced commercial typography;
-- perfect 2x2 title grids;
-- shiny black texture, gray shading, gradients, embossed text, or vector-clean edges;
-- adding pinyin, English, captions, labels, dates, numbers, or decorative text.
-
-## Three-Part Typography Workflow
-
-For publishable `带字版`, use separate assets instead of asking one image generation to solve everything.
-
-Pass 1: poster base
-
-- Preserve one real object as the photographic anchor.
-- Remove busy surroundings and create a warm white field.
-- Reserve a large empty title area in the upper-left or left side.
-- Do not generate the title in this pass.
-- Add at most one quiet micro-worker near the object if it helps the object story.
-
-Pass 2: title construction reference sheet
-
-- Build a model-facing private reference sheet with `scripts/build_title_reference_sheet.py` when local cutouts exist.
-- Pass the exact target title to the script only so it can build the right number of off-grid layout boxes; keep the semantic text in the prompt. Include one to three `lettering/title-blocks` specimens, several `lettering/glyphs` specimens, and several `lettering/strokes` specimens.
-- Use the sheet to show glyph skeleton defects: off-grid character boxes, non-parallel strokes, bad joins, blunt ends, uneven scale, and awkward internal spaces.
-- Do not render the target title as a system font on the model-facing sheet. Keep exact semantic text in the prompt, and let the sheet control only layout pressure, bad spacing, and malformed stroke references.
-- Do not treat the construction sheet as the final poster. It is a control image for the title layer.
-
-Pass 3: title layer
-
-- Generate or create only the black handwritten title on white or transparent background.
-- Use the title construction reference sheet as the primary control reference.
-- Iterate the title layer by itself until the handwriting works.
-- Composite it onto the poster base without changing the object. Use `scripts/composite_title_layer.py` when working with local image files.
-
-Direct all-in-one generation is allowed only for fast drafts. If the user complains that the type is not close enough, switch to the title construction sheet and title-layer workflow.
-
-For maximum fidelity, use one of these title sources, in order:
-
-1. a real hand-drawn title written by the user or designer and scanned/photo-cleaned;
-2. a custom title asset drawn by tracing the private lettering references;
-3. a title-only image-model generation guided by a construction reference sheet;
-4. an all-in-one poster generation.
-
-Treat options 3 and 4 as draft quality when exact resemblance matters. The reference style depends on malformed character skeletons, not just texture, so a normal generated Chinese glyph will usually remain too font-like.
-
-## Building a Title Construction Sheet
-
-When `private-assets/reference-cutouts/asset-index.json` exists, create a construction sheet before writing the title-layer prompt. The default sheet is for the image model: no English labels and no target characters drawn in a standard font.
-
-```bash
-python scripts/build_title_reference_sheet.py \
-  --title "泰香开饭" \
-  --out private-assets/reference-cutouts/title_reference_sheet_taixiang.png \
-  --title-ref title_tiramisu_heavy_black \
-  --title-ref title_salty_cheese_crooked
-```
-
-Use the output as a reference image for the title layer. The sheet is most useful when the requested title does not share characters with the references. It tells the model which defects to transfer: clumsy skeletons, broken joins, childlike spacing, and non-font rhythm.
-
-Use `--human-labels` only when you need to inspect which assets were selected. Do not feed that labeled inspection sheet to an image model, because labels and standard-font target text can leak into the generated title.
-
-If the generated title still looks like a font, keep the poster base and regenerate only the title layer with the same construction sheet. If the exact characters still fail, ask for or create a real handwritten title asset and composite it locally.
-
-## Four-Character Title Layout
-
-For four Chinese characters, do not make a neat square grid. Use two loose vertical columns:
-
-```text
-泰     开
-香     饭
-```
-
-But the columns should be staggered:
-
-- left column can start higher than right column;
-- right column can lean inward or drift down;
-- the lower characters should not align perfectly;
-- column spacing can be slightly too tight or too loose;
-- character boxes should feel hand-placed, not mechanically distributed.
-
-When prompting, describe it as:
-
-```text
-two loose vertical handwritten columns, not a 2x2 table; left column contains "泰" above "香", right column contains "开" above "饭"; each character has a different size, tilt, x-position, and vertical gap
-```
-
-## Title-Layer Prompt Template
-
-Use this when generating the title separately. If a title construction sheet exists, pass it as the main reference.
-
-```text
-Create a black handwritten Chinese title layer on a plain white background.
-
-Text, exact:
-[title]
-
-Layout:
-[layout description]
-
-Style:
-Use [title construction sheet path] as the primary reference. Match the glyph skeleton defects and stroke anatomy from the sheet. The title should look hand-written by an untrained child using a blunt black marker: flat black ink, crooked independent strokes, uneven character sizes, broken blunt stroke ends, awkward spacing, off-grid placement, bad joins, collapsed radicals, imperfect but readable Chinese. The characters should not sit on a perfect grid. No calligraphy, no brush font, no sans-serif font, no rounded cute handwriting, no elegant tapering, no clean font skeleton.
-
-Output constraints:
-Only the black title on white or transparent background. No object, no poster, no doodle figure, no logo, no extra text, no pinyin, no English, no numbers, no watermark.
-```
-
-## Poster-Base Prompt Template
-
-Use this for Pass 1 of `带字版`.
-
-```text
-Create the typography-led poster base from the uploaded photo, but do not render the title yet.
-
-Preserve [object] as a real photographic object with its original texture, color, shadow, and material. Remove the busy source background and place the object on a clean warm white poster field. Reserve a large empty title area in the upper-left or left side for a later black handwritten Chinese title layer. Keep the object smaller than the title area, usually lower-right or lower-middle, with enough white space around it.
-
-Add at most one tiny primitive black-line micro-worker near the object, not in the title area. The worker should be functional, faceless, and hand-drawn, not cute.
-
-No generated title, no captions, no logo, no watermark, no random text. If source object contains real printed text, keep it only if it is part of the photographed object and not visually dominant.
-```
-
-## Composition Corrections For Food Photos
-
-Food-table photos usually start too busy. In `带字版`, do not keep the whole table scene.
-
-Use this structure:
-
-- choose one plate or one food basket, not the whole meal;
-- object height: 30 percent to 42 percent of canvas;
-- title block: 35 percent to 50 percent of canvas width and 30 percent to 45 percent of canvas height;
-- title must have clean breathing room and should not touch the object;
-- crop away competing plates, table texture, and decorative clutter unless one piece is the chosen object;
-- if the plate rim has strong printed letters, keep them only inside the photographic object and prevent them from competing with the black title.
-
-## Revision Prompts
-
-## Local Compositing
-
-After creating a poster base and a title-layer image, composite them locally:
-
-```bash
-python scripts/composite_title_layer.py \
-  --base poster-base.png \
-  --title title-layer.png \
-  --out poster-final.jpg \
-  --x 8% \
-  --y 7% \
-  --width 42%
-```
-
-Use a larger width for title-first posters and a smaller width when the object needs more room. The script thresholds dark pixels from the title layer, turns them into black ink, removes the white background, and places the result on the base.
-
-If the title looks like a font:
-
-```text
-Regenerate only the title layer. Do not change the poster base. Use the title construction sheet more literally for glyph skeleton defects: off-grid character boxes, uneven character sizes, crooked independent strokes, ugly joins, blunt broken ends, non-parallel horizontals, collapsed radicals, no calligraphy rhythm, no clean font skeleton.
-```
-
-If the layout is too much like a menu photo:
-
-```text
-Regenerate only the poster base. Remove the table-scene feeling. Keep one real food object lower-right or lower-middle, reserve a large blank title field on the left, and reduce all other objects and decorative clutter.
-```

package/skills/heytea-doodle-poster-hosted/references/mixed-media-style-guide.md

@@ -1,101 +0,0 @@
-# Mixed-Media HEYTEA-Inspired Style Guide
-
-Read this reference for desktop-pet candidate generation. Also read it for poster work when the requested direction is crayon, collage, pencil, or mixed media.
-
-This guide describes a visual study of user-supplied references. It captures transferable drawing grammar, not official brand assets. Never reproduce source logos, complete campaign copy, packaging labels, or the drinking-head mark unless the user supplied them, has the right to use them, and explicitly asks for them.
-
-## Core visual system
-
-The reference language is not a uniform cartoon outline or a fully packed crayon drawing. It is a loose collage assembled from four unequal layers:
-
-1. thin black or muted-green pen lines establish structure;
-2. heterogeneous dry media—light colored-pencil scumbling, childlike wax-crayon loops or blocks, and occasional dry pastel—supply most color;
-3. pale wash or translucent blocks indicate liquid, clear plastic, light, or atmosphere;
-4. occasional photographic ingredient fragments create a deliberate real-versus-drawn collision.
-
-Do not force every layer into every image. Two or three layers are enough, but the result must retain visible material contrast.
-
-## Line grammar
-
-- Mix thin-to-medium pen with strategically heavier marker passages. For `flavor-monster`, the body and source-derived protrusions require one semantically complete heavy black structural stroke system, while the face and 2+2 limbs remain thinner. At any cross-section the boundary reads as one mark, never a parallel band, braid, rope, or outline-of-an-outline. Every important stroke must change pressure or thickness within itself; a complete boundary is not permission for constant-width monoline.
-- Every structural stroke must visibly wander along its full length. Cup walls, rims, bases, arms, legs, and facial dashes should use short flat spots, shallow dents, drifting curvature, pressure jumps, imperfect corners, dry breaks, overshoots, or mismatched joins. For a flavor-monster boundary, allow only one or two short local retraces and keep continuous parallel travel below roughly ten percent of the complete path. Breaking ruler-straight lines into fragments, bending a smooth Bézier path, or adding surface jitter to a clean vector skeleton does not create the target hand feel.
-- Allow open joins, overshooting terminals, and color that misses the outline by a few pixels.
-- Keep secondary marks sparse. One loose curl, broken arc, crooked underline, or short motion dash is more faithful than dense decoration.
-- Make the geometry itself structurally naive: unequal wall slopes, tilted rims, off-center openings, shifted lids, inconsistent perspective, and crooked bases. A clean template with surface wobble is not enough.
-
-The target is an observed hand making decisions in real time. Smooth Bézier curves, mechanically rounded corners, sticker borders, identical line weight, whole-loop mechanical doubling, braided or rope-like tracks, repeated beaded stamps, and clean vector paths with a wobble filter fail this style.
-
-## Color and material grammar
-
-- Begin with the photographed subject's one to three identifying colors.
-- Favor food-native, low-to-medium-saturation hues: tea green, yellow-green, cream, pale yellow, fruit red, and restrained brown.
-- Assign media by material region instead of applying one texture to the whole character. `source-faithful` may keep loose colored pencil, wax crayon, graphite, or occasional pale wash according to source material. `flavor-monster` uses dry wax crayon as the dominant color tool.
-- For `flavor-monster`, wax crayon uses a small number of broad blunt side-of-crayon sweeps with matte granular deposits, broken pressure, ragged ends, and visible paper tooth. Follow the approved material layer map: establish a continuous primary matrix, embed inclusions or deposits within it, keep surface layers spatially distinct, and overlap broad dry-wax passes to create intermediate hues and depth. Paper-white breathing space is optional and must not sever the mapped hierarchy. Slight misses or crossings at the black boundary are allowed when the boundary remains readable.
-- Broad rubbed pigment or pale wash is an occasional accent, not the dominant default. Oil-paint, oil-pastel smear, gouache mass, wet watercolor bloom, dense repeated tiny strokes, crosshatching, airbrushed gradients, soft plastic shading, glossy highlights, and uniform digital noise are forbidden.
-- Black is structural rather than dominant. White or warm white remains an active material, not merely unused background.
-- Photographic fragments, when used, should be small ingredient accents. Do not turn the entire desktop-pet body back into a photo.
-
-## Composition and character grammar
-
-- Preserve generous negative space. Review boards should feel like objects placed on paper, not a filled sticker sheet.
-- Cuteness comes from awkward scale, a clear verb, and the relationship between body and prop—not from large eyes or a detailed face.
-- Human figures are faceless, back-facing, or reduced to functional contours. Limbs bend simply and may be anatomically naive.
-- For a `source-faithful` desktop pet, the photographed object remains the body. For a `flavor-monster` desktop pet, the approved creature remains the body; the photo supplies color, ingredient, material, and body topology as static visual identity DNA, while its motion verb is reserved for later behavior design. Every flavor-monster face mark and all four limbs use sparse black doodle lines.
-- Keep the two arms and two legs short, open, black, and slightly clumsy. Let pose, tilt, squash, crumbs, bubbles, steam, or wobble carry emotion.
-
-## Model-facing reference boards
-
-Inspect these boards before writing an identity or action prompt:
-
-- `examples/desktop-pet/pink-green-flavor-monster-v3/preview.png`: public, repo-created positive style anchor for HEYTEA-like cuteness and handmade looseness. It is not an anatomy or outline template: never copy its body topology, protrusion placement, facial-mark type/count/layout, limb attachment points, lengths, proportions, resting pose, feature coordinates, or outline treatment into a new identity.
-- `private-assets/reference-cutouts/desktop-pet-style/contact_sheet_single_pass_rough_line_v1.png`: primary anatomy-free `flavor-monster` black-line board, assembled only from non-lettering HEYTEA source cutouts. Its isolated heavy contour curves and lighter arm/leg curves control one-pass path construction, pressure jumps, blunt terminals, crooked travel, dry breaks, bad joins, and restrained local retracing. It supplies no silhouette, face, limb skeleton, object, text, or character anatomy.
-- `private-assets/reference-cutouts/desktop-pet-style/contact_sheet_crayon_layer_v1.png`: color-stage-only board for broad blunt dry wax-crayon, paper tooth, and optional breathing space. It supplies no color-zone structure or required gap count. Do not pass it to flavor-monster black-line-master generation, and do not copy its line fragments.
-- `private-assets/reference-cutouts/desktop-pet-style/contact_sheet_dry_media_wavy_line_v4.png`: primary line-and-color reference for continuously meandering strokes, light colored-pencil scumbling, childlike wax-crayon loops, chunky crayon marks, and paper tooth.
-- `private-assets/reference-cutouts/desktop-pet-style/contact_sheet_smudged_paint_structure_v3.png`: secondary reference for fading edges, paper intrusion, crooked rims, unequal walls, and shifted lids; do not let its broad smears dominate the whole character.
-- `private-assets/reference-cutouts/desktop-pet-style/contact_sheet_figure_actions_v2.png`: faceless action figures, primitive props, gesture economy, and object-to-figure scale.
-- `private-assets/reference-cutouts/desktop-pet-style/contact_sheet_thin_stroke_anatomy_v2.png`: thin black/green line weight, awkward glyph structure, open joins, overshoots, retracing, and dry breaks.
-- `private-assets/reference-cutouts/desktop-pet-style/contact_sheet_mixed_media_objects_v2.png`: analysis-only example of an over-hatched failure. Never pass it as a positive identity-generation reference.
-
-For `flavor-monster` black-line generation, pass the uploaded photo, public pink-green example, and the single-pass rough-line board with a separate declared role for each. Construct the face from the approved written face brief only. After the line master passes, use only the locked master, crayon-layer board, and dry-media board for color. Exclude the lettering-derived thin-stroke board from flavor-monster identity generation, use the figure-action board only during motion generation, use the smudged-paint board only for focused material repair, and keep the mixed-media-object failure board outside every positive input. No reference board may supply silhouette, face, limb skeleton, object, text, layout, logo, packaging, or anatomical coordinates.
-
-For `source-faithful`, continue choosing the same positive controls according to the photographed object's construction needs. When these study boards ship with the open-source Skill, identify them as non-official reference assets and keep them outside the CC BY grant for original examples.
-
-When a generated identity still has a clean cartoon framework, borrow line anatomy—not figure identity—from primary poster-mode `figures/full-poses/*` and `lettering/strokes/*` cutouts. `flavor-monster` always uses a black-line master followed by a line-locked crayon pass. Poster assets control segment construction, blunt ends, line-weight jumps, and bad joins; they must not introduce workers, lettering, props, or logos into the pet.
-
-## Desktop-pet translation
-
-For every `source-faithful` three-candidate board and every `flavor-monster` single identity master:
-
-- use mode-specific hesitant pen; for `flavor-monster`, enclose the body and source-derived protrusions with a heavy irregular black boundary and keep the face plus 2+2 limbs thinner but pressure-varying;
-- preserve mode-specific material layering; for `flavor-monster`, use broad blunt dry wax-crayon to preserve the primary matrix, embedded inclusions or deposits, surface layers, overlaps, optional breathing space, and selected missed edges rather than transparent wash, peer color blocks, or one equal-density texture;
-- use a warm-white review background with enough separation to inspect edge quality.
-
-For `source-faithful`:
-
-- preserve the source silhouette, major internal divisions, identifying colors, and must-preserve object details;
-- vary face placement, limb proportions, structural crookedness, and temperament only;
-- reject animal costumes, monster substitution, or loss of object recognition.
-
-For `flavor-monster`, obtain all identity decisions from `desktop-pet-character-modes.md`. This guide controls only their visual execution: heavy irregular boundary, thinner line features, broad blunt dry wax-crayon, paper tooth, optional breathing space, and the assigned reference-board stages.
-
-For animation frames, preserve the same stroke logic even when poses change. Runtime interpolation may move or squash the complete drawing, but it must not replace hand-drawn silhouette-changing key poses.
-
-## Rejection checklist
-
-Reject and regenerate when any of these dominate:
-
-- smooth, equal-width, sticker-like, whole-loop mechanically doubled, braided, rope-like, outlined-band, or repeated-stamp black outline;
-- smooth vector geometry or sticker border;
-- airbrushed gradient, glossy 3D volume, or plastic shading;
-- transparent marker, watercolor, pastel or oil-pastel smear, smooth digital fill, or dense uniform crayon covering every region with the same mark;
-- paper tooth or digital speckle presented as "white space" while color still fills the body edge to edge;
-- smooth constant-width face or limb curves, including clean Bézier paths with only superficial jitter;
-- more than one or two local retraces, or continuous parallel travel covering roughly ten percent or more of the complete flavor-monster boundary;
-- oil-paint, oil-pastel, gouache, or wet-smear appearance dominating the character;
-- ruler-straight segments, even when separated by deliberate gaps;
-- structurally clean geometry with a superficial wobble filter;
-- generic chibi face as the main source of cuteness;
-- excessive props, decorations, text, logos, or campaign layout;
-- photographic desktop-pet body with doodles merely pasted on top.
-- `source-faithful` output that replaces the source object with a generic creature.
-- `flavor-monster` body missing its heavy irregular structural boundary, or using a smooth equal-width/sticker-like boundary instead.

package/skills/heytea-doodle-poster-hosted/references/monster-poster-workflow.md

@@ -1,56 +0,0 @@
-# Approved-Monster Fusion Poster Workflow
-
-Read this reference only after a `flavor-monster` identity has received explicit user approval and the user chooses a fusion poster.
-
-## Required inputs
-
-Use both of these inputs and keep their roles separate:
-
-1. The original uploaded photo is the sole source of the product. Preserve its photographic texture, silhouette, proportions, construction, liquid or food regions, translucency, condensation, highlights, and material cues. Background cleanup and subject cutout are allowed; redesigning, repainting, or cartoonizing the product is not.
-2. The approved canonical monster is the sole identity source for the character. Lock its body topology and proportions, face layout, exactly two arms and two legs, attachment points, palette, dry-media treatment, flavor-DNA marks, and line character. Pose may change to support the interaction; identity may not.
-
-Do not regenerate a new monster from the product photo during poster production. If the approved identity artifact is missing or ambiguous, stop and ask the user to provide or reconfirm it.
-
-## Interaction concept
-
-Choose one clear interaction verb per poster: `投`, `扶`, `抱`, `探`, `推`, `尝`, or another single primitive action the product can physically support.
-
-- Make the product and monster visibly affect one another through contact, overlap, gaze, weight, direction, or a short action arc.
-- Keep the action legible without explanatory copy.
-- Use one main interaction, not a collage of unrelated gags.
-- Remove background props that do not support the action, including loose straws or duplicated packaging when they compete with the subject.
-- Preserve generous intentional white space and keep the product as the photographic anchor.
-
-## Text poster
-
-Read `style-guide.md` and `lettering-guide.md`. Build a composition designed for typography from the start:
-
-1. Generate the poster base with the photographic product, approved monster, interaction, and reserved title zone, but no generated title glyphs.
-2. Build the title construction sheet with `../scripts/build_title_reference_sheet.py` when applicable.
-3. Create the title as a separate transparent layer and combine it with `../scripts/composite_title_layer.py` when applicable.
-
-The title should follow the requested crooked handmade hierarchy and the source composition's planned negative space. It must not cover defining product structure, the monster's face, its 2+2 limb inventory, or the contact point of the interaction.
-
-## No-text poster
-
-Forbid all words, letters, numbers, captions, labels, pseudo-glyphs, signatures, and watermarks. Do not reserve an empty title box. Let the product scale, monster pose, contact point, and one interaction verb complete the story.
-
-## Two-version rule
-
-When the user requests both versions, design two independent compositions. Change the product placement, monster staging, balance of white space, or interaction framing so each version is resolved for its own narrative. Do not create one shared base and merely toggle the title layer.
-
-## Review gate
-
-Reject or revise when any of these occur:
-
-- the photographed product's structure, material, or proportions drift;
-- the product becomes illustrated, repainted, or replaced by a generated lookalike;
-- the monster's body, face, limb count, attachment points, colors, material, or flavor marks drift from the approved canonical identity;
-- pose adaptation silently becomes identity redesign;
-- more than one interaction competes for attention, or no physical relationship is readable;
-- the text version relies on image-model typography instead of the separate title workflow;
-- the no-text version contains any text-like marks;
-- the two versions reuse the same composition with text switched on or off;
-- runner availability or installation status blocks this static poster branch.
-
-Use `evaluation.md` for the final combined poster review.