Cards To Images

Goal

Turn an approved card package into local image assets that a person can inspect and upload.

Generate actual files, verify them one by one, and keep platform upload and public-URL preparation behind visible handoff states.

Inputs

Collect:

• approved card package from long-to-cards or a compatible source
• delivery_mode=publish_ready
• platform_profile: xiaohongshu, wechat-inline, zhihu-idea, or generic
• ordered card ids, copy, visual direction, source anchors, and alt text
• visual system, brand constraints, and supplied assets when available
• output directory
• existing content_state
• generation_backend=auto by default for generative or mixed work
• an existing chatgpt-image-handoff run when external images are returning to the pipeline
• asset_url_policy: local by default or public when stable URLs are required

Route back to long-to-cards when the sequence, copy, source anchors, or image order is incomplete. Route back to research or review when a card depends on an unsupported claim.

Platform Profiles

Use current reviewed platform requirements when they exist. Record any deliberate override in the manifest.

Workflow

1. Validate The Card Package

Confirm:

• card ids and image order are unique and complete
• every card has one main idea
• copy, source anchor, visual direction, and alt text are present
• the cover, caption, CTA, and final card agree
• claims and evidence have already passed their required gates

Do not render an unstable package merely to create a preview that may be mistaken for a final asset.

2. Select A Render Strategy

Choose one strategy per visual system:

• deterministic: use canvas, HTML, SVG, diagrams, or another exact renderer for text-heavy cards, flows, checklists, tables, and factual structures
• generative: use image generation for conceptual or atmospheric illustration; add exact labels with a deterministic layer when needed
• mixed: combine generated visuals with deterministic typography and layout

Prefer deterministic rendering when Chinese text, Skill names, arrows, counts, or source-backed relationships must be exact.

3. Resolve Generative Work

For generative or mixed strategies, route the image jobs through resilient-imagegen with generation_backend=auto.

• use built-in ImageGen when healthy
• use chatgpt-image-handoff when Computer Use or manual ChatGPT handoff is selected
• import every external result under the stable card filename before layout QA
• keep exact Chinese, Skill names, counts, arrows, and source-backed relationships in a deterministic layer
• use a fully deterministic renderer when external ChatGPT transmission is declined

Do not treat a prompt pack or a generator UI result as a rendered card. The branch rejoins only after a local file exists.

4. Render Real Files

Create each ordered PNG or JPG inside the reviewed project directory. Preserve stable filenames such as 01-cover.png, 02-context.png, and 03-method.png.

Treat delivery_mode=publish_ready as approval to produce local files. Pause before paid generation, editing sensitive supplied assets, or any action that needs a new external permission.

5. Inspect And Revise Every Image

Open every rendered image and check:

• exact text, names, numbers, and page order
• legibility at expected mobile size
• overflow, clipping, unintended line breaks, and safe margins
• correct dimensions and aspect ratio
• consistent colors, typography, spacing, and series markers
• visual alignment with the source claim and card purpose
• absence of invented metrics, interfaces, quotes, logos, or outcomes

Revise failed cards and inspect them again. A contact sheet can support overview review, but it does not replace per-image inspection.

6. Write cards-manifest

Record one row per output with:

• card_id
• platform_profile
• source_card
• local_path
• format, width, and height
• alt_text
• render_strategy
• generation_backend
• handoff_run_id, when an external handoff was used
• generation_status
• visual_qa_status
• human_confirmation
• asset_url_policy
• r2_state
• public_url

Use these states:

• generation_status: not_started, planned, generated, or failed
• visual_qa_status: pending, passed, or needs_revision
• human_confirmation: pending, approved, or rejected
• r2_state: not_planned, planned, confirmed, uploaded, or verified

7. Hand Off

• With asset_url_policy=local, return the local images and manifest for human confirmation and direct platform upload.
• With asset_url_policy=public, route the reviewed local references to md-img-r2 plan mode. Require explicit confirmation before upload or URL rewrite.
• When a chatgpt-image-handoff pack exists, preserve its run_id, imported-file checksum, and QA result in the cards manifest.
• Route the assembled package to wenchang-publish-check final mode. Keep human_confirmation=pending until the user actually approves the images.

Output Format

# Card Image Delivery

## Render Summary
- Platform profile:
- Card package:
- Output directory:
- Image count:
- Render strategy:
- Generation backend:
- Handoff run:
- Asset URL policy:

## Visual QA
| Card | File | Dimensions | Generation | Visual QA | Human confirmation |
| --- | --- | --- | --- | --- | --- |

## Manifest
- File:
- R2 state:
- Public URL state:

## Handoff
- Next skill:
- User decision needed:
- Blockers:

Boundaries

• Do not claim completion when only prompts, layouts, or card copy exist.
• Do not mark visual QA as passed without opening every generated image.
• Do not silently change approved card copy while rendering.
• Do not upload to R2, rewrite Markdown URLs, or publish to a platform without explicit confirmation.
• Do not expose private paths, credentials, account sessions, or unreleased source material in public outputs.

作者入口

• 微信公众号：AI生命克劳德
• X：@yangchao228
• GitHub：yangchao228