snowmanzhuang/yixueaiganhuo-ppt/yixueaiganhuo-ppt

yixueAIganhuo-PPT

yixueAIganhuo-PPT is now the 唯一完整 skill for the end-to-end workflow:

medical sources
  -> 医学上游 worker
  -> ppt_plan.json and slide prompts
  -> 页面生成 worker via editppt image
  -> page worker editable reconstruction
  -> editppt run finalize
  -> final editable PPTX

The 统一父 agent in this skill is a scheduler. It creates the task directory, maintains global state, dispatches workers, controls concurrency, collects structured results, and runs deterministic editppt commands. It does not personally absorb full PDF text, every Figure detail, every prompt, OCR details, or page-level manifest.json decisions into its context.

Parent context budget: keep the parent thread limited to routing decisions, worker dispatch, state paths, validation summaries, and final reporting. Do not perform medical PPT planning in the parent thread. The parent must dispatch the 医学上游 worker before reading full source text and must keep structured artifacts only in its own context.

Do not ask the user to approve sub-agent spawning. Once the skill is chosen and prerequisites are satisfied, sub-agent dispatch is part of the workflow. Do not ask for a reply such as a permission phrase; proceed with worker dispatch and report progress.

References

• references/single-skill-prd.md: product and architecture baseline for the merged single-skill workflow.
• references/workflow-overview.md: upstream medical planning DAG.
• references/prompt-key-selection.md: style key selection and per-slide prompt construction.
• references/style_selector/: local HTML style selector page and 001-019 template/example preview assets.
• references/provider-config.md: unified editppt image backend contract.
• references/medical-integrated-workflow.md: parent orchestration, worker boundaries, concurrency, incremental page intake, and failure handling.
• references/cli-helper.md: editppt install check and command manual.
• references/page-decision-tree.md: page-worker object-source decisions.
• references/manifest-schema.md: editppt run/page JSON contracts.
• scripts/start-style-selector.py: launches the local style selector page from the skill bundle.
• scripts/classify-task-route.py: lightweight helper for initial task-route classification.
• scripts/validate-slide-prompt.py: full-authoring rich GPT Image 2 prompt gate before slide-image worker dispatch.
• scripts/validate-slide-ready.py: full-authoring preflight gate before downstream conversion commands.
• prompts/medical-upstream-worker.md: prompt template for medical planning workers.
• prompts/medical-slide-image-worker.md: prompt template for single-slide image generation workers.
• prompts/page-worker.md: prompt template for downstream editable page workers.

Retained Capabilities

The skill keeps the original medical creation value:

• source reading and concise medical summary
• PDF/Figure inventory and evidence Figure selection
• target slide structure and narrative planning
• ppt_plan.json
• Chinese speaker notes
• page dependencies
• per-slide GPT Image prompt writing

The skill also embeds the downstream editable reconstruction value:

• editppt CLI
• page worker dispatch
• manifest.json
• OCR/text hints
• foreground image asset reconstruction
• PPT native element reconstruction
• page validation
• final deck assembly through editppt run finalize

Deprecated Components

The old yixue text-overlay editable path is deprecated. Do not use legacy OCR/image-repair PPTX scripts as the current execution path. The authoritative editable path is now the embedded editppt page-worker workflow.

Task Routing

Before creating downstream jobs, inspect the user request and input files, then write task_route into task_state.json.

Recommended helper:

python yixueAIganhuo-PPT/scripts/classify-task-route.py \
  --user-request "<original user request>" \
  --input-summary "<brief first-page/overview/OCR/metadata notes>" \
  <input paths...>

Use the helper output as structured guidance, then apply the rules below. If a PDF preview clearly shows title/authors/abstract/body/references, choose FULL_AUTHORING_FROM_SOURCE even when the user says "转 PPTX".

Do not treat a research/material PDF as an existing visual slide deck merely because it is a PDF.

Full Authoring Hard Gate

For FULL_AUTHORING_FROM_SOURCE, the final PPTX is invalid unless every slide passes the complete chain:

1. prompts/slideNN.txt follows the rich prompt schema from references/prompt-key-selection.md and passes:

python yixueAIganhuo-PPT/scripts/validate-slide-prompt.py \
  --prompt <task>/prompts/slideNN.txt \
  --slide-number <N>

1. slides/slideNN.png is generated by editppt image generate/edit/batch through the 页面生成 worker.
2. slide_results/slideNN.json has status: "IMAGE_READY", passed: true, and image_backend.tool_call showing editppt image generate/edit/batch.
3. scripts/validate-slide-ready.py passes for every generated slide before editppt prepare or editppt run add-page.
4. Every page is reconstructed by a page worker, recorded by editppt run record, and assembled only by editppt run finalize.

Do not use @oai/artifact-tool, python-pptx, HTML/SVG screenshots, local PIL composition, direct native PPT generation, legacy yixue scripts, or any presentation/PPTX authoring shortcut as a substitute final path. If workers or image backend are unavailable, stop and report the blocker; do not downgrade to a simpler PPTX path.

Preflight Questions Gate

Before dispatching the 医学上游 worker for FULL_AUTHORING_FROM_SOURCE, ask a concise preflight question block unless all required values are already known.

Before showing the block, start the bundled style selector server from this skill and use the printed 点击打开：... URL as {style_selector_url}. Do not automatically open the browser; present the URL in the question block so the user can click it when needed:

python3 yixueAIganhuo-PPT/scripts/start-style-selector.py --background

If the local server cannot start, use the printed 本地文件兜底：... file URL instead. The launcher resolves paths relative to its own script location, so it works when the whole skill folder is copied to another computer or installed under another username.

The block must start with: "请先回答下面几个问题："

Use numbered emoji labels only:

请先回答下面几个问题：
 
1️⃣ 页数：这次生成多少页 PPT？
 
2️⃣ 风格：请打开风格选择页，选 001-019 后回复序号。
{style_selector_url}
不选则默认：001 通用提示词风格。

If no PaddleOCR-VL token is detected, append the third question:

3️⃣ OCR：未检测到 PaddleOCR-VL token。
需要更准的文字识别，请打开这里获取 token：
https://aistudio.baidu.com/account/accessToken
然后回复 token；也可以回复“跳过”继续。

If a PaddleOCR-VL token is already detected, do not show the OCR prompt; ask only questions 1️⃣ and 2️⃣. The OCR token question is optional and must not block generation if the user chooses to skip.

The page count remains a hard gate. Ask the user for the target slide count before dispatching the 医学上游 worker unless the request already gives an explicit count. Do not default to any page count. Do not offer preset page-count choices. Record the confirmed number in task_state.json as target_slide_count_confirmed and pass it to the upstream worker as TARGET_SLIDE_COUNT.

The style selector is optional. If the user gives a valid selector (001-019, a matching style name fragment, or a style JSON filename/path), record it in task_state.json as style_selector and pass it to the upstream worker as STYLE_SELECTOR. If the user does not choose a style, record and pass STYLE_SELECTOR=001.

If the user gives a range, ask for one concrete number. If the user refuses to decide, stop and ask for a numeric page count; do not silently choose a count and do not present preset options.

Parent Workflow

1. Create one task directory, for example output/yixueAIganhuo-PPT/<run_id>/.
2. Classify the task with the Task Routing rules and write task_route to task_state.json.
3. If task_route is DIRECT_VISUAL_CONVERSION, follow Direct Visual Conversion Workflow below and skip the full-authoring steps 4-15.
4. Write task_state.json with run phase, worker ids, per-slide status, paths, retries, concurrency slots, task_route, style_selector, and optional style_selector_url.
5. Dispatch the 医学上游 worker using prompts/medical-upstream-worker.md. The parent passes TARGET_SLIDE_COUNT only after target_slide_count_confirmed=true, and passes STYLE_SELECTOR after the user chooses a style or the parent defaults it to 001; it does not create the PPT outline or slide prompts itself.
6. Receive only structured upstream artifacts:

• source_inventory.json
• figure_inventory.json
• plans/ppt_plan.json
• prompts/slideNN.txt

1. Validate that ppt_plan.json has one entry per target slide, continuous slide numbers, speaker notes, asset bindings, prompt files, and page dependencies.
2. Validate every slide prompt before dispatching 页面生成 workers:

python yixueAIganhuo-PPT/scripts/validate-slide-prompt.py \
  --prompt <task>/prompts/slideNN.txt \
  --slide-number <N>

Do not dispatch a 页面生成 worker with a summary-style prompt, a prompt that only says "medical academic style", or a prompt that lacks layout geometry, evidence/asset bindings, negative constraints, and continuity rules.

1. Configure or verify the unified image backend:

editppt setup
editppt doctor

Use editppt config only when API fallback credentials are needed.

1. Dispatch 页面生成 worker jobs:

• slide 1 first
• slide 2 after slide 1 exists, with slide 1 as style reference
• slide 3+ after slide 2 exists, with slide 1 and slide 2 as style references

1. Every 页面生成 worker must call editppt image generate/edit/batch. It writes slides/slideNN.png and slide_results/slideNN.json.
2. Before the first generated slide enters downstream conversion, run the real slide image preflight:

python yixueAIganhuo-PPT/scripts/validate-slide-ready.py \
  --slide <task>/slides/slide01.png \
  --result <task>/slide_results/slide01.json