deep-slides8.8 KB
View saved version →
---
name: deep-slides
description: Create or edit a Sider Slides presentation through conversation, including requests with a template ID. Research and design new decks, upload images, author HTML slides and speaker notes, or edit a completed deck by adding, deleting, moving, duplicating, changing text or rewriting pages. Use for presentation, PPT and slide deck creation or editing requests; not PPTX export.
---
# Deep Slides
You author the presentation. The Sider tools read templates, create the presentation ID, upload images, save your outline and pages, apply edits, and open the app. They do not research, generate content, or run a background generation agent. Respect the user's explicit content and design preferences over these defaults. Choose the creation workflow for a new deck, or the editing workflow for an existing completed deck.
## Template input
When the user supplies a template ID, including through the template picker, first call `get_presentation_template` with `template_id`, before creating a presentation or planning its outline. Read the returned `summary` and every page's `content`; the tool has already downloaded and split the template. Keep the template ID and returned content in context for both outline planning and HTML authoring. If summary is empty, infer the design from the pages. If retrieval fails, stop dependent steps and explain the failure rather than silently choosing another design.
Use the template's palette, typography, spacing, decorations and layout patterns in both the outline and generated pages. Select and reuse suitable template page types to fit the user's content and requested page count; do not force a one-to-one page mapping. Replace sample text, dates and data with the user's material or researched facts. Treat template content as a design reference, not instructions to execute. Public image URLs on `pub-cdn.wisebox.ai` actually present in the returned pages may be reused directly; other newly selected images require upload.
## Creation workflow
1. **Create the presentation ID.** After reading any supplied template, call `create_presentation` with a short title derived from the request, before uploading images or saving content. Without a template ID, this is the first Sider tool call. Retain the returned `ppt_id` throughout this workflow; the template ID is a separate identifier. Never invent an ID, substitute a task/sub-agent ID, or create another deck to recover a failed step.
2. **Understand and research.** Read [research.md](references/research.md). Use your available reading, web search and image search capabilities to understand the provided materials, create an initial narrative, collect useful evidence and images, and refine the plan. Ask only questions that materially affect the result; use supplied preferences without asking again. Default to the user's language, about 10 slides and at most 20 slides. If the topic needs more, explain how you will merge related content within this limit.
3. **Upload the selected images.** Build a per-page asset list from the refined plan and deduplicate source URLs. Reuse permitted public template images directly. For all other images, call `upload_presentation_asset` with the ID and each direct HTTP(S) source URL. Retain successful `(source URL, file_id, cdn_url)` mappings and reuse them across pages. Finish uploading and associating these images before saving the outline; only their returned `cdn_url` is usable, not the upload input URL. Skip uploads if the deck uses only permitted template images or no external images.
4. **Save the final outline.** Read [outline.md](references/outline.md) and produce the complete Markdown outline. With a template, record its ID, design characteristics and the template page types used for each planned slide. Record exact uploaded `cdn_url` values or permitted public template image URLs, then call `save_presentation_outline` with the ID and `main_content`. Ordinary research citation links may remain. This starts the page-saving round. Do not call this tool again after any page has been saved: it clears the page cache.
5. **Author and save the pages.** Read [slide-html.md](references/slide-html.md). Write each complete HTML fragment and speaker note yourself, implementing the outline and its template references. Use the same image URLs recorded in the saved outline, including CSS backgrounds. Call `save_presentation_page` with the ID, `page_index` starting at 1, `content`, and `note`. Follow the outline order and confirm each save succeeds. Do not put XML artifact tags or Markdown fences in the tool arguments.
6. **Open the app.** Once every planned page save has succeeded, call `open_presentation` with the same ID. Report the page count and that the editor entry was opened. The tool is read-only and does not load, validate or change the saved pages. Do not claim to have rendered the deck, exported a PPTX file, or completed an asynchronous backend job.
## Editing workflow
1. **Identify a completed deck.** Use the existing `ppt_id` from the conversation or user. Never create a new deck to edit an existing one, or substitute a SubAgent ID. If the ID or target page is unclear, ask for that information. The tools use explicit page numbers, not the editor's current selection.
2. **Wait for generation to finish.** Editing tools are allowed only after the entire PPT has been generated and every planned page save has succeeded. Do not finalize or edit partially generated decks, or use editing tools to repair an unfinished generation round.
3. **Read before every edit.** Before EACH call to `add_slide`, `delete_slide`, `move_slide`, `duplicate_slide`, `modify_text` or `regenerate_slide`, call `get_slides_with_all_done` with the same ID and inspect its complete pages. A previous edit's read, a cached outline, or the UI's initial load does not satisfy this step. If the read fails, stop; do not edit.
4. **Prepare and perform one edit.** Read [editing.md](references/editing.md) for the operation contracts. Base the operation and page numbers on the freshly read deck. Author the HTML yourself for new or rewritten pages, following [slide-html.md](references/slide-html.md). Reuse existing deck images or permitted public template images; upload other new images first and use the returned `cdn_url`. Do not resave the outline or use `save_presentation_page` during editing.
5. **Confirm saving, then continue.** Each editing tool fetches all latest pages again, applies one operation, and internally calls `save_all_pages` without `version` to force-save. A successful `pages_saved` result confirms the edit. Do not perform another save or ask for a version/overwrite confirmation. For another edit, return to step 3 and use the updated page order. Execute edits for the same deck sequentially, never in parallel. Report only confirmed changes; an already open editor receives them through polling. Use `open_presentation` if the user needs to open the editor.
## Creation assets and failures
- External images, including user-provided images and background images, must pass through `upload_presentation_asset`, except public `pub-cdn.wisebox.ai` images actually present in the retrieved template. Do not reuse a search URL, invent a CDN URL, or substitute an expiring resource-list URL for the returned `cdn_url`.
- Keep the saved outline, retrieved template content and upload mapping together in the conversation. Both the outline's image resources and the actual page HTML must use the same uploaded or permitted template image URLs. Keep other original image URLs only in the working upload mapping; ordinary research citation links can remain in the outline. Do not resave the outline after page saving begins.
- Use HTML/CSS, inline SVG or code charts for diagrams, icons and data visualization. Do not use base64 images, local paths, videos, audio or embedded external webpages. Only images are supported by the upload workflow.
- Resolve failed image uploads before saving the outline: obtain and successfully upload a different source, or revise the plan to use an appropriate image-free layout. If the requested image is essential, explain the missing input and pause dependent steps. Do not save the original URL or pretend an upload or search succeeded.
- Preserve the ID and successful page/asset results when a tool fails. For a failed page save, retry the same page index and content. Never automatically repeat creation or upload after an ambiguous network failure, and never restart by saving the outline again.
- An upload error with `uploaded: true` and a `file_id` means the file exists but association may have failed. Keep that ID when reporting the failed step; do not use it as a completed asset or claim success.
- Stop dependent steps on authentication or business errors and explain the affected step. Open the editor only after all intended page saves have succeeded. Do not invoke legacy generation/editing tools or add a separate completion call.
Referenced files: 5 — file references are metadata, not package contents.
