AI Content Images

Image Card Series Generator

Break down complex content into eye-catching image card series with multiple style options.

User Input Tools

在 DSH 中需要确认时调用 ask_user_question。如果一次要确认多项，合并在同一次调用中；只在该工具不可用时，才退回编号文字问题。

Image Generation Tools

本案例固定调用 DSH Plugin 工具 generate_content_image。该工具通过受限的本机 codex exec --ephemeral 桥接 $imagegen；这是本案例插件补充的能力，不是 DSH 原生生图能力。调用前必须先把每张图的完整提示写入独立文件，再把本次 project_dir 和相对路径交给工具。

⛔ Never substitute SVG, HTML, canvas, or other code-based rendering for raster image generation. Codex imagegen's own description says it should be used "when the output should be a bitmap asset rather than repo-native code or vector." If you cannot resolve a raster backend via step 3, fall through to step 4 and ask the user — do not silently emit SVG, write inline <svg> markup, or produce HTML/CSS art as a substitute. This applies even if the article/section seems "diagram-like": the consumer skill calling this rule has already decided that a raster image is what it needs.

⛔ Never repair rendered text by painting over a generated bitmap. Do not use ImageMagick, Pillow, Canvas, SVG, HTML/CSS, OCR scripts, or any other programmatic overlay to cover, rewrite, erase, stroke, or replace titles, body copy, tags, or any other text inside an already generated image card. If text is wrong or unclear, regenerate from a corrected prompt, switch to a layout with less on-card text, or ask the user which imperfect candidate to keep.

Prompt file requirement (hard): write each image's full, final prompt to a standalone file under prompts/ (naming: NN-{type}-[slug].md) BEFORE invoking any backend. The file is the reproducibility record and lets you switch backends without regenerating prompts.

生成后调用 inspect_content_series，传入同一个 project_dir 和期望张数，检查本次新系列的 PNG 签名、尺寸、数量与提示文件数量。

Batch Generation Policy

After every prompt file for the current generation group has been saved and verified, generate images in batches by default.

在 DSH 中按顺序调用 generate_content_image。插件为保护本地 Codex 会话强制并发数为 1；“批量”仍表示确认全部提示文件后连续生成，不改变下方的首图锚点链。

Rules:

• Honor the image-1 anchor chain: generate image 1 first, then batch images 2+ using image 1 as the reference.
• Never start a batch until every selected prompt file for that batch exists on disk.
• Retry failed items once without regenerating successful items.
• Do not use subagents merely to parallelize image rendering. Use subagents only for separate prompt iteration or creative exploration.

Confirmation Policy

Default behavior: confirm before generation.

• Treat explicit skill invocation, a file path, matched signals/presets, and EXTEND.md defaults as recommendation inputs only. None of them authorizes skipping confirmation.
• Do not start Step 3 until the user completes Step 2.
• Skip confirmation only when the current request explicitly says to do so, for example: --yes, "直接生成", "不用确认", "跳过确认", "按默认出图", or equivalent wording.
• If confirmation is skipped explicitly, state the assumed strategy / style / layout / palette / count / backend in the next user-facing update before generating.

Language

Respond in the user's language across questions, progress, errors, and completion summary. Keep technical tokens (style names, file paths, code) in English.

Options

Dimensions

Three independent knobs combine freely:

Example: --style notion --layout dense makes an intellectual knowledge card; add --palette macaron to soften the colors without changing notion's rendering rules. A --preset is a shorthand for style + layout (+ optional palette).

Palette behavior: no --palette → style's built-in colors; --palette <name> → overrides colors only, rendering rules unchanged. Some styles declare a default_palette (e.g., sketch-notes defaults to macaron).

Styles (12)

Per-style specifications: references/presets/<style>.md.

Layouts (8)

Layout specs: references/elements/canvas.md.

Palettes (optional override)

Replaces the style's colors while keeping rendering rules (line treatment, textures) intact.

Palette specs: references/palettes/<palette>.md.

Presets (style + layout shortcuts)

Quick-start combos, grouped by scenario. Use --preset <name> or recommend during Step 2.

Knowledge & Learning:

Lifestyle & Sharing:

Impact & Opinion:

Trend & Entertainment:

Poster & Editorial:

Full prompt-fragment definitions: references/style-presets.md.

Auto-Selection

Match content signals to the best combo. First row whose keywords appear wins; fall back to cute-share if nothing matches.

Style × Layout Matrix

Compatibility scores (✓✓ highly recommended, ✓ works well, ✗ avoid). Use when the user picks a non-default combo and you want to flag a poor match.

Outline Strategies

Three differentiated approaches — each produces a structurally different outline. The workflow recommends one; Path C generates all three and lets the user choose.

Reference Images

User-supplied refs are separate from the internal "image-1 as anchor" chain (Step 3) — they layer on top of it.

Intake: via --ref <files...> or paths pasted in conversation.

• File path → copy to refs/NN-ref-{slug}.{ext}
• Pasted with no path → ask for the path, or extract style traits as a text fallback

Usage modes (per reference):

Record refs in each affected card's prompt frontmatter:

references:
  - ref_id: 01
    filename: 01-ref-brand.png
    usage: direct

At generation time: verify files exist. Image 1 with usage: direct + backend that accepts refs → pass via the backend's ref parameter (becomes the chain anchor). Images 2+ keep using image-1 as --ref per Step 3 — do NOT re-stack user refs on top (avoids conflicting signals). For style/palette, embed extracted traits in every prompt.

File Layout

generated/{topic-slug}/
├── source-{slug}.{ext}
├── analysis.md
├── outline-strategy-{a,b,c}.md    # Path C only
├── outline.md
├── prompts/NN-{type}-{slug}.md
├── output/NN-{type}-{slug}.png
└── refs/                          # only if --ref used

Slug: 2-4 words, kebab-case. "AI 工具推荐" → ai-tools-recommend. On collision, append -YYYYMMDD-HHMMSS.

新生成硬约束：每次都从用户本轮输入建立新的 generated/{topic-slug}/。不得读取案例根目录的 input/deepseek-harness-article/ 或 output/deepseek-harness/ 来代替分析、提示或生成结果；它们只是教师展示案例。

Backup rule (applies throughout): before overwriting any file — source, outline, prompt, image — rename the existing one to <name>-backup-YYYYMMDD-HHMMSS.<ext>. This protects user edits.

Workflow

- [ ] Step 0: Load EXTEND.md ⛔ BLOCKING (interactive only)
- [ ] Step 1: Analyze content → analysis.md
- [ ] Step 2: Smart Confirm ⚠️ REQUIRED (Path A / B / C)
- [ ] Step 3: Generate images
- [ ] Step 4: Completion report

Step 0: Load EXTEND.md ⛔ BLOCKING

Check these paths in order; first hit wins:

• Found → read, parse, print a summary (style / layout / watermark / language), continue.
• Not found + interactive → run first-time setup (see references/config/first-time-setup.md) and save before anything else. Do NOT analyze content or ask style questions until preferences exist — this keeps first-run behavior predictable.
• Not found + --yes → skip setup, use built-in defaults (no watermark, style/layout auto-selected, language from content). Do not prompt, do not create EXTEND.md.

EXTEND.md keys: watermark, preferred style/layout, custom style definitions, language preference, preferred image backend, generation batch size. Schema: references/config/preferences-schema.md.

Step 1: Analyze Content → analysis.md

1. Save the source (backup rule applies if source.md exists).
2. Run the deep analysis in references/workflows/analysis-framework.md: content type, hook potential, audience, engagement signals, visual opportunity map, swipe flow.
3. Detect source language, pick recommended image count (2-10).
4. Auto-recommend strategy + style + layout + palette using the Auto-Selection table above.
5. Write everything to analysis.md.

Step 2: Smart Confirm ⚠️ REQUIRED

Hard gate: this step is mandatory per the Confirmation Policy — Step 3 cannot start until the user confirms here (or explicitly opts out with --yes / equivalent wording in the current request).

Goal: present the auto-recommended plan and let the user confirm or adjust. Skip this step entirely under --yes — proceed with Path A using the analysis and any CLI overrides.

Display summary before asking:

📋 内容分析
  主题：[topic] | 类型：[content_type]
  要点：[key points]
  受众：[audience]

🎨 推荐方案（自动匹配）
  策略：[A/B/C] [name]（[reason]）
  风格：[style] · 布局：[layout] · 配色：[palette or 默认] · 预设：[preset]
  图片：[N]张（封面+[N-2]内容+结尾）
  元素：[background] / [decorations] / [emphasis]

Then ask one question — three paths. Verbatim option copy: references/confirmation.md.

Path A — Quick confirm (trust auto-recommendation): generate a single outline using the recommended strategy + style → save to outline.md → Step 3.

Path B — Customize: ask five questions (strategy/style, layout, palette, count, optional notes) with the recommendation pre-filled — blanks keep the recommendation. Generate one outline with the user's choices → outline.md → Step 3. See references/confirmation.md.

Path C — Detailed mode: two sub-confirmations.

• Step 2a — Content understanding: ask selling points (multi-select), audience, style preference (authentic / professional / aesthetic / auto), optional context. Update analysis.md.
• Step 2b — Three outline variants: generate outline-strategy-a.md, outline-strategy-b.md, outline-strategy-c.md. Each MUST have a different structure AND a different recommended style — include style_reason in the frontmatter. Page-count heuristic: A ~4-6, B ~3-5, C ~3-4. Template: references/workflows/outline-template.md; frontmatter example in references/confirmation.md.
• Step 2c — Selection: ask three questions (outline A/B/C/Combined, style, visual elements). Save selected/merged outline to outline.md → Step 3.

Step 3: Generate Images

With confirmed outline + style + layout + palette:

Visual consistency — image-1 anchor chain: character / mascot / color rendering drifts between calls unless you anchor them. Generate image 1 (cover) first WITHOUT --ref, then pass image 1 as --ref to every subsequent image. This is the single most important consistency trick for this skill — don't skip it even if the backend also supports a session ID.

Generation flow:

1. Write the full prompt for every image to prompts/NN-{type}-{slug}.md in the user's preferred language (backup rule applies), then verify all selected prompt files exist.
2. Generate image 1 first without --ref; backup rule applies to the PNG file. This establishes the anchor.
3. Build a task list for images 2+ using image 1 as --ref <path-to-image-01.png>.
4. 按 ## Batch Generation Policy 顺序生成 images 2+；每张都调用一次 generate_content_image，参数中的 project_dir 始终为本次 generated/{topic-slug}，prompt_file 为相对该目录的 prompts/...，参考图为相对该目录的 output/01-....png。
5. Report progress after each completed image. On failure, retry only the failed item once from the same saved prompt file.
6. 全部生成后调用 inspect_content_series({ project_dir, expected_count })，只验收本次新系列；不使用教师案例的 manifest。

Watermark (if enabled in EXTEND.md): append to the generation prompt:

Include a subtle watermark "[content]" positioned at [position].
The watermark should be legible but not distracting.

See references/config/watermark-guide.md.

Backend selection: 使用 generate_content_image。Prompt files MUST exist before invoking the tool. 该桥接的边界见 references/codex-imagegen.md。

Session ID (if the backend supports --sessionId): use cards-{topic-slug}-{timestamp} for every image; combined with the ref chain this gives maximum consistency.

Step 4: Completion Report

Image Card Series Complete!

Topic: [topic]
Mode: [Quick / Custom / Detailed]
Strategy: [A/B/C/Combined]
Style: [name]
Palette: [name or "default"]
Layout: [name or "varies"]
Location: [directory]
Images: N total

✓ analysis.md
✓ outline.md
✓ outline-strategy-a/b/c.md (detailed mode only)

- 01-cover-[slug].png ✓ Cover (sparse)
- 02-content-[slug].png ✓ Content (balanced)
- ...
- NN-ending-[slug].png ✓ Ending (sparse)

Content Breakdown Principles

For the style × layout compatibility matrix, see the Style × Layout Matrix above.

Image Modification

Always update the prompt file before regenerating — it's the source of truth and makes changes reproducible.

Text correction policy:

• If a card's title, body copy, tags, or any other rendered text is misspelled, garbled, hard to read, or visually weak, do not patch the bitmap with code.
• For text-correction regenerations, write a new prompt file and a new output path so the flawed candidate is preserved for comparison.
• Post-processing is limited to crop, resize, compression, or format conversion that does not alter text or the main composition.

References

Notes

• Auto-retry once on generation failure before reporting an error.
• For sensitive public figures, use stylized cartoon alternatives.
• Smart Confirm (Step 2) is required; Detailed mode adds a second confirmation (2a + 2c).

Changing Preferences

EXTEND.md lives at the first matching path listed in Step 0. Three ways to change it:

• Edit directly — open EXTEND.md and change fields. Full schema: references/config/preferences-schema.md.
• Reconfigure interactively — delete EXTEND.md (or ask "reconfigure ai-content-images preferences" / "重新配置"). The next run re-triggers first-time setup.
• Common one-line edits:
  • preferred_image_backend: dsh-content-imagegen — 本案例固定的 DSH Plugin bridge。
  • generation_batch_size: 1 — 本案例强制串行，保护本地 Codex 会话。
  • preferred_style: notion, preferred_layout: dense, preferred_palette: macaron, language: zh.
  • watermark.enabled: true + watermark.content: "@handle" — add a watermark.