Generate real PNG/JPG, and save a reusable Prompt, source boundaries, manifest, and acceptance results for each generation. This skill is not responsible for article layout, platform publishing, account login, or fabricating facts out of thin air.
Customer-readable description
What this skill can do
Supports six delivery types: article covers/summary cards, Cornell notes, visual metaphor posters, skill library promo single images/carousels, plugin icons, and brand Logo systems. Customers only need to provide the source, purpose, platform, aspect ratio, and text that must appear verbatim; when reference images are available, specify whether they are style references, composition references, or editing targets.
How customers use it
- First clarify
image_type, purpose, and output form; omitted combination axes are assumed by the Agent. - When the request is vague, first show the L0 support catalog and let the customer choose
presetorfamily; do not silently apply a good-morning, font-mask, or any aesthetic preset. - After selection, load only the matching references, compile the complete Prompt, call imagegen, and execute
view_imageand the corresponding quality gate.
When requirements are unclear: ask first, do not generate
When the customer only says "make a nice-looking image" or "help me put together a poster" or does not give a purpose/output form, first run the L0 catalog command, then clarify with the minimum questions below; do not guess a preset, and do not generate a sample image first:
我先确认一下需求。我们目前支持：
1. 文章封面 / 研究小结卡
2. 康奈尔笔记信息图
3. 视觉隐喻海报
4. 技能库宣传单图 / 小红书轮播
5. 插件或应用图标
6. 品牌 Logo 系统（图形标、字标、组合和变体）
7. 公开 X Prompt Deck → image 技能进化导入
请回复：
- 用途：封面、小结、笔记、海报、宣传卡、轮播、插件图标还是品牌 Logo？
- 来源：文章/仓库/X Prompt Deck，还是只给一个主题？
- 风格：从支持目录选一个 family，或描述你想要的视觉感觉？
- 版式：比例、平台、张数；有没有必须逐字出现的标题、数字、URL、字标或 Logo？
When the customer answers only some questions, ask only about the missing fields that would change routing; for secondary axes such as aspect ratio and text strategy, the Agent provides a confirmable default value. Only when the customer explicitly says "generate directly" and the purpose, source, and output form are already sufficiently certain may this round of questions be skipped.
Dependencies and installation
Requires a callable imagegen and view_image; when missing, stop the corresponding stage and do not substitute code-based drawing. Plugin installation, configuration files, and provider login state are handled according to the host documentation; this skill does not save API keys, cookies, or tokens.
Progressive selection and loading (must be followed)
L0: Support catalog
On the first request, read only template-registry.yml and prompt-composition-index.yml 's support_catalog. The current catalog contains 7 delivery templates, 11 Prompt families, 4 Logo variants, and 1 external prompt evolution import route.
python3 scripts/resolve_prompt_composition.py --list-supported
Show the customer the counts, templates, family summaries, and external routes; before the customer has selected a purpose/output form/family, do not load all long Prompts.
L1: Select family
Confirm family, purpose, information structure, and output form. For example presentation_grid + deck_series + 16:9、celebration_ceremony + campaign_pack, or brand_identity + logo_system + 1:1。
L2: Combination axes
Load only one family、information_structure、visual_mechanism、aesthetic_system、text_strategy and the corresponding entries for model_adapter , as well as family catalog; then read combination block execution contract. For missing axes, provide one confirmable default value; do not treat adjectives as an already-compiled Prompt.
L3: Generation and acceptance
Read the selected template, source facts, and quality gate, write the seven canonical blocks to disk, then call imagegen; actually open the bitmap and verify text/aspect ratio/facts, and if it fails, regenerate with a new version.
Routing index
- Ordinary article image: read input contract、template registry and combination index。
- Research cover/summary:
editorial_research_minimal; body summary:editorial_summary_card; Cornell notes:cornell_notes。 - Delivery selection support
social_card | carousel | icon | logo; social graphics usesocial_skill_catalog, plugin icons useplugin_icon, brand Logo usesbrand_logo. For the facts and two-stage rules of single-image Moments posts or Xiaohongshu carousels, see Social Card Contract。 - Research covers use
editorial_research_minimal; for plugin icons and brand Logo imagegen, only direction drafts are provided, and the final version must be deterministically redrawn as vectors. The brand Logo additionally reads Logo System Prompt。 - X profile import:
source=x-profile-exportwhen, first read X Import Contract and X Prompt Evolution Contract, run:
python3 scripts/import_x_profile_prompt_deck.py \
--input <x-profile-run>/image-prompts.yml \
--output <image-output>/x-profile-evolution
When an inventory and gap analysis against current capabilities is needed, run:
python3 scripts/audit_x_profile_prompt_deck.py \
--input <x-profile-run>/image-prompts.yml \
--output <image-output>/gap-report.md
This step is not copying the Prompt: it verifies run manifest.yml, window/filter conditions, status ID, source evidence, and canonical blocks, then breaks out base_visual_system、topic_seasoning、series_variables、render_plan. GPT2 is retained only as a source model label and does not become an image skill dependency.
Input and Prompt Compilation
For complete fields, see Input Contract. The minimum input must include source、image_type、preset/family, purpose, ratio, output directory, and verbatim text; for the Logo, the brand name, graphic concept, wordmark strategy, color variants, and vector final-version requirements must also be provided; for X import, the source URL, status ID, selection filters, source_prompt, and seven blocks must also be retained.
Each generation is written to disk in the following order, without copying entire sections from the case library:
composition_axes: record the delivery template, family, information structure, visual mechanism, aesthetic system, text strategy, model adaptation, and batch strategy.source_grounding: the facts allowed into the image, the fields that must not be filled in, and missing information.primary_task: one primary visual task, without mixing the cover, summary, and infographic into one image.composition_and_layout、visual_style_and_materials: subject, space, reading order, material, light, and color roles.exact_text、aspect_and_output、constraints_and_avoid: verbatim text, ratio/format, and verifiable prohibitions.
Covers and social cards must distinguish the base visual system, topical accents, series variables, and factual layer; within the same series, only 2–4 variables change per image.GPT2 x 早安 x 字体蒙版is a combination query, not a new preset.
Execution Loop
- Confirm: parse L0–L2 and record assumptions for unspecified axes.
- Evidence gathering: only
social_skill_catalogrunbuild_social_catalog_facts.py; repository-recommended/key skills must also read the repository README and the key skill SKILL.md. - Compile: create
prompts/NN-*.md, and write all seven blocks and source evidence; do not generate images until it is written to disk. - Generate: ordinary templates use the Agent built-in imagegen by default;
plugin_iconandbrand_logo's second stage may use deterministic SVG/vector scripts, and must not use HTML, canvas, Pillow, or screenshots to pass off as the final version. - Acceptance: according to Quality Gate , actually
view_image; social cards then runvalidate_social_catalog_delivery.py, and check facts, OCR, QR codes, thumbnails, and semantic density. - Regenerate: only change one main issue, use a new prompt and new filenames, regenerate consecutively at most twice; do not overwrite versions that have already passed.
Delivery Directory and Data Boundaries
When not specified output_dir , you must first run:
python3 scripts/resolve_output_dir.py --source <source> --json
Adopt the parsing result and record it in the manifest output_dir_origin. It is forbidden to treat cwd, the skill repository, the vault root, or <topic>-delivery-<date> as the default delivery root; do not save caches or unselected candidates in the delivery directory. For the complete directory, configuration, temporary data, and receipts, see Delivery Contract。
A normal delivery includes at least prompts/, the final PNG/JPG, and manifest.yml; the two-stage template additionally includes a fact list and reproducible synthesis sources. An X evolution delivery additionally includes evolution.yml、series-index.yml、bibles/ and recompilation prompts/。
Private Information and Intermediate Data
Configuration goes only in a private skill-specific root; intermediate candidates go in an OS temporary run directory and are cleaned up afterward. Provider caches, cookies, tokens, unrelated private files, and absolute temporary paths do not enter the public delivery. For the complete rules, see Delivery Contract。
Hard Boundaries
- Do not forge QR codes, installation commands, project URLs, existing brand logos, numbers, or personal identities;
brand_logois used only for new brand identities explicitly requested by the client, and must not impersonate a third party or an existing brand. - Do not treat a script returning 0 as an image passing; a prompt compiling successfully also does not mean the bitmap passes.
- Do not read unrelated private directories, and do not write provider caches, cookies, tokens, or temporary paths into the public delivery.
- When imagegen or
view_imageis missing, normal images and Logo direction drafts stop with a clear explanation; the second Logo stage can execute deterministic SVG only when an approved path, vector specifications, and visual acceptance conditions are all complete, and must not use code output to impersonate a direction draft.
Verification
python3 scripts/generate_skill_catalog.py --check
python3 scripts/generate_expert_manifest.py --check
python3 scripts/audit_skills.py
python3 -m unittest discover -s tests -p 'test_*.py'
Forward testing must inspect each real bitmap one by one; X evolution testing must also check base_visual_system、topic_seasoning、series_variables and render_plan, and must not merely check whether the YAML exists.
Receipt after completion
完成：<一句话结果>
路由：image_type=<值> preset=<值> family=<值> aspect=<值>
产出：Prompt=<路径> 图片=<路径/尺寸> manifest=<路径>
验证：事实=<通过/不适用> 文字=<通过/失败> 比例=<通过/失败> view_image=<通过/失败>
问题与下一步：<无；或列出未通过项>