Easyslides

Academic PPTX Skill

Generate and edit academic/research PPTX presentations with native editable DrawingML shapes.

EasySlides is a project-backed skill. This file is the agent entrypoint and task router; real deck generation depends on the full repository: scripts/, workflows/, templates/, references/, tests, and local runtime dependencies. Installing only SKILL.md is enough for routing guidance, but not enough to generate or validate PPTX files.

Read ARCHITECTURE.md for the layer model and capability paths. Read INSTALL.md for minimal skill, full local runtime, and developer installation modes.

Route Authority

Default onboarding is a conversation, not a browser workflow. Ask targeted native popup questions, one consequential question at a time, adapting to each answer until the brief is executable. Skip known facts; resolve conflicting requirements; then summarize and proceed. Follow skills/easyslides-clarify/SKILL.md. Open HTML guides or browser confirmation pages only when explicitly requested.

Before choosing an implementation path, read workflows/routing.md. It owns deterministic route selection for main generation, raw PPTX template fill, native PPTX enhancement, reusable template creation, slide-image reconstruction, audio, preview, and chart verification. The workflow registry is workflows/index.md.

Before selecting a route, invoke the clarification gate in skills/easyslides-clarify/SKILL.md. If the request leaves more than one reasonable interpretation that would affect the route, story, page count, template, visible wording, or visual fidelity, ask the user to choose from explicit options. Do not write a deck plan or generate slide files while a blocking choice remains unanswered.

Mandatory production-scheme choice

For every new or regenerated PPT, an unspecified production scheme is a blocking choice even when the topic, template and content are otherwise clear. MUST ask the user to choose and wait before planning slides or generating images/SVG/PPTX. Follow workflows/clarification-gate.md for the exact three options and examples. No automatic default. Reuse only the user's explicit choice for the same task.

Mandatory reconstruction-mode choice

In both choice rounds, MUST show Token 消耗 and 耗时 levels for every option before the user chooses. Explain that levels are relative estimates adjusted for scope and complexity, not exact usage or promised time; disclose image generation separately. Follow the cost and time disclosure in the clarification workflow.

For image reconstruction, confirm 全图矢量重建 (full_vector) or 保留复杂配图 (preserve_complex_images) through workflows/clarification-gate.md before execution. No automatic default, even for “快点做”; reuse only an explicit choice in the same task. Both modes require native PPT text boxes. Preserve one source line in one text box, merging OCR fragments and using text runs for mixed styling; keep independent labels and table cells separate. Full-vector mode needs approval before any raster exception. Partial rebuilds apply this within selected regions.

Image-reference templates

Read the lightweight registry metadata first; load only the user's selected reference PNG. Do not preload the whole image library or batch-OCR templates. Use compressed previews for browsing. The installed library shares the project directory via a local junction; do not duplicate assets when syncing the plugin.

For the image-based schemes, select visual references from templates/image_references/registry.json or user-supplied images, independently of the native editable template library. Follow the image-reference routing in workflows/routing.md; reference PPT-to-contact-sheet work does not distill a native template or generate a new deck.

Chinese wording defaults

Use “垂听” instead of “聆听” in assistant-authored Chinese PPT text by default. For example, write “感谢垂听，敬请讨论”. This user wording preference applies to all three production schemes, including slide copy, imagegen prompts, SVG/native text, speaker notes and final wording checks. It is a user preference, not a claim about universal language usage. A later explicit user instruction can override it; do not rewrite verbatim source quotations or archived files.

Core Capabilities

1. Create from scratch: Source content → SVG pages → DrawingML shapes → editable PPTX
2. Edit existing PPTX: Unpack → edit XML → validate → repack
3. HTML/JSX authoring path: Use HTML/JSX as an experimental upstream authoring layer, then normalize into EasySlides-compatible SVG or shape IR before the existing DrawingML backend
4. Academic scenario/template system: Scenario-first academic planning with template-extensible output. The official template set is academic_general, academic_scqa, defense_leftnav, defense_topnav, literature_minimal, nsfc_defense, and thu_speech; templates/template_policy.json is authoritative.
5. 71 visualization templates: Charts, infographics, diagrams, strategic frameworks, and tables
6. 11,635 icons: Six SVG icon families for consistent academic and business visuals, with lucide preferred for new generic icons and emoji replacement
7. Template Asset Bank: Convert many real PPTX templates into exact-reuse slide modules for manual-template-substitution quality
8. PPT Master Page Recipe Library: whole-page SVG layout archetypes with fixed regions, text slots, diversity rules, and strict SVG slot measurement
9. Card Component Library: 13 fixed-size card styles plus PPT Master-style visual recipes with slot capacity contracts, agent selection rules, prompt skeletons, and PPTX preview export
10. PPT Master compatibility mode: Strict serial Strategist/Executor workflow, hand-written SVG pages, native DrawingML export, rebuildable backups, and gate checks via scripts/ppt_master_pipeline.py
11. Native PPTX template fill route: Raw PowerPoint templates route to workflows/template-fill-pptx.md instead of the SVG pipeline.
12. Native PPTX enhancement route: Finished decks route to workflows/native-enhance-pptx.md for notes/audio/timing/transition patches without visible-slide regeneration.
13. Confirmation page: Package deck-plan assumptions with scripts/confirm_ui.py before visual execution when the user wants a reviewable confirmation page.
14. Visual review package: Render or reuse slide PNGs through scripts/visual_review.py to produce a review manifest, local HTML page, and contact sheet.
15. Brand presets: Register reusable palette/typography/logo inputs with scripts/create_brand.py under templates/brands/.
16. Clarification gate: Resolve result-affecting ambiguity through explicit user choices before route selection and execution.

Backend-Centered Architecture

EasySlides has one production backend: normalized SVG/shape IR converted to editable DrawingML/PPTX by the local scripts/svg_to_pptx/ pipeline. Multiple authoring frontends may feed that backend, but do not introduce a second production PPTX backend unless a dedicated spike proves better editable output, Office compatibility, and lower maintenance cost.

Path A: Create from Scratch (SVG → DrawingML → PPTX)

When no template PPTX exists, use the SVG-to-DrawingML pipeline:

Source Content → Project Init → Strategist (Deck Plan + Design Spec) → SVG Generation → Quality Check → Export PPTX

Key advantage: Every SVG element becomes an editable DrawingML shape — text is selectable, colors are changeable, shapes are movable.

Path B: Edit Existing PPTX (XML Unpack/Edit/Repack)

When a template PPTX exists:

Template PPTX → Unpack XML → Edit Content → Clean Orphans → Validate → Repack PPTX

Key advantage: Preserves all template formatting, animations, and layout structure.

If the user asks to improve the look of an existing PPTX while preserving slide count, order, and visible wording, route through workflows/beautify-pptx.md. That route currently supports inspection plus a conservative native theme-color patch; stronger layout repair must not silently restructure the deck through the main generation path.

Path C: HTML/JSX Authoring (Experimental Upstream)

When the source is a rendered HTML page, dashboard, report, or SVG-heavy technical slide, use HTML/JSX only as an authoring and measurement layer:

HTML/JSX source -> browser measurement / component declarations -> normalized SVG or shape IR -> EasySlides validation -> DrawingML export

Key advantage: HTML/JSX is easier for LLMs to author for complex layouts, component reuse, dashboards, and architecture diagrams. The output still flows through the EasySlides checker and DrawingML backend.

Boundary: @artifact-kit/pptxgenjs-jsx and html-to-pptx-skill may be used as references or in isolated spikes, but they are not main-path dependencies of EasySlides. Prefer integrating their DOM measurement and JSX authoring ideas upstream of the existing SVG/DrawingML backend.

Use workflows/html-jsx-authoring.md when evaluating this path.

Path D: PPT Master Compatibility Mode

When the user explicitly asks to reproduce PPT Master, use the PPT Master workflow, or prioritize PPT Master-style SVG execution, follow workflows/ppt-master-compat.md.

This mode keeps EasySlides' local converter stack but mirrors PPT Master's operating contract:

Source -> project init -> Strategist design_spec/spec_lock -> optional images
-> Executor hand-writes SVG pages sequentially -> svg_quality_checker
-> validate_svg_text_slots -> notes/total.md -> ppt_master_pipeline export
-> editable PPTX

The Executor must write SVG pages directly into svg_output/; do not create a batch script that generates all pages. Use references/ppt-master-compat.md for execution rules, templates/page_layouts/ppt-master-page-recipes-manual.md for whole-page layout selection, and templates/cards/assembly-manual.md for nested card selection.

For a large library of PPTX templates, prefer the Template Asset Bank harness:

PPTX templates → pptx_template_import.py workspaces → template_asset_bank.json → exact slide-module reuse

Use workflows/template-asset-bank.md and scripts/template_asset_bank.py when the goal is to mimic manual template substitution: fixed geometry, fixed decorative structure, and only text/image/chart data replaced.

Path E: Slide Image Reconstruction

When the source of truth is a slide screenshot, exported slide PNG, or AI mockup image, use workflows/slide-image-to-editable-pptx.md. This is an upstream analysis and QA contract for the existing EasySlides backend: first classify every visible source-image element into Layer A visual assets, Layer B native structure, or Layer C editable text; then assemble through SVG/shape IR and DrawingML export.

This path explicitly forbids full-slide screenshot backgrounds, baked text in image assets, and dirty rectangular crops. Validate the Phase 1 element inventory with scripts/slide_image_inventory.py, then validate the final deck with scripts/validate_image_reconstruction_pptx.py alongside the normal text layout and render-diff gates.

Use scripts/project_manager.py init <name> --kind slide_image_reconstruction and scripts/image_reconstruction_pipeline.py init/qa as the standard project facade for this path. Default QA mode is pixel-strict: editable text, native structure, asset safety, provenance and source-render differences block delivery. --mode faithful-practical is diagnostic only; a visual failure sets delivery_ready=false and a failing CLI exit code. Always inspect actual renders. For generated complete slides, follow the host-call and generation-manifest contract in workflows/slide-image-to-editable-pptx.md. Attach the selected template image and original evidence to the actual ImageGen call. Reconstruct each approved output as its sole visual source; do not reuse native page shells. Complex raster illustrations that become ugly as vectors should use preserve_source_frame; closed/circular source assets should use masked source assets with clipping checks. The image route also carries a shared alignment contract: Layer C source boxes map to absolute 1280x720 canvas coordinates, editable text stays outside transformed parent groups, and center-locked labels are checked against the actual native PPTX frame after export. Cover, content, navigation, and ending-page chrome use the same contract, so a visually hidden SVG offset cannot pass the final gate.

Path F: Native PPTX Template Fill

When the user provides a raw PowerPoint template and asks to fill it with new material or a new topic, route to workflows/template-fill-pptx.md. This path preserves the source deck as a native slide library and patches cloned slides directly through OOXML. Do not convert the raw PPTX template into SVG for a one-off fill request.

Path G: Native PPTX Enhancement

When the user provides a finished PPTX and asks only for speaker notes, narration audio, auto-advance timings, or page transitions, route to workflows/native-enhance-pptx.md. This path is append-oriented and must not regenerate or rewrite visible slide content.

Path H: Visual Review and Brand Presets

When the user asks for a confirmation page or checklist before visual execution, route to workflows/confirm-ui.md. When the user asks for a visual self-check, shareable preview, or review page for an existing deck, route to workflows/visual-review.md. When the user asks to create, register, or inspect a reusable brand palette/logo, route to workflows/create-brand.md.

Academic Scenario-First Template Contract

Use scenario first, template second for academic work. In plain terms: scenario first, template second. The academic scenario defines the argument structure, evidence obligations, audience state, and page roles; the template route defines visual containers, geometry, palette, chrome, and reusable slide modules.

• A template route is not a scenario: template route is not a scenario. Do not infer "thesis defense" just because a defense-looking template is selected, and do not infer "literature report" just because a literature-style shell is available.
• The built-in scenario ids are seed profiles for routing, not a closed taxonomy. If the user's academic use case is not an exact match, select the nearest profile and record scenario_variant in deck_plan.json and design_spec.md.
• Preserve the selected profile's hard and required rules, then explicitly note which recommended or relaxable rules change for the scenario_variant.
• When no matching visual template exists, do not force a defense or literature-report template. Use free academic design, a general academic pack, a domain pack, or a user-provided template path while keeping source traceability, citation retention, text fit, and PPTX deliverability intact.
• For every academic task, read references/academic-orchestration.md before intake and planning. Audience-State-Transfer and source-backed narrative are shared planning capabilities across templates and production schemes; SCQA is an optional spine. Respect user outlines and approved image pages.

Source Material Policy

Before planning or generating a deck, classify the user's input:

• No supplied source materials: establish whether the user will supply materials or wants agent research. An explicit research request already resolves this choice. For authorized research, run workflows/topic-research.md. You may gather web text and download relevant openly licensed images as PPT assets, then import the research document and image folder as source materials with provenance.
• Provided mature source materials: if the user supplies a journal paper, thesis/dissertation, mature report, existing deck, or similar file/URL with substantive text, figures, tables, or captions, treat that material as the source of truth. Build from its claims, structure, figures, tables, captions, and user-provided assets. Do not replace extracted figures, invent substitute evidence visuals, change the source claims, or add outside material unless the user explicitly asks; any added generic background/icon must remain decorative and never stand in for source evidence.

Path A: Create from Scratch

Step 1: Requirements & Project Init

python scripts/project_manager.py init <project_name> --format ppt169
python scripts/project_manager.py import-sources <path> <files...> --move

Supported source formats: PDF, DOCX, XLSX, PPTX, URL, Markdown.

MinerU PDF Preprocessing: When importing PDFs, the system automatically tries MinerU for structural extraction before falling back to PyMuPDF. MinerU provides richer output: Markdown with figure/table identification, layout JSON, and extracted images — enabling better PPTX generation with automatic figure extraction.

To enable MinerU: set MINERU_API_TOKEN environment variable or create a .mineru_token file in the project root.

Paper-Report Intake

For single-paper report decks, run the paper-report intake after importing sources and before Strategist writes the final deck plan:

python scripts/paper_intake.py <project_path> --json

The intake reads <project_path>/sources/, Markdown converted from the PDF, MinerU/PyMuPDF-side manifests when present, and <project_path>/images/. It drafts deck_plan.json with a source_map for the main paper plus extracted figures, then validates the draft with scripts/deck_plan_contract.py. Strategist should treat the output as a traceable starting point: verify the paper title, figure captions, claims, and slide roles before writing the final design_spec.md and spec_lock.md.

The intake also records source-located evidence blocks in each page's content_contract (conclusion, evidence, explanation) and summarizes content_quality (status, evidence count, source-text length, and material types). The component planner uses these signals to select an argument stack, evidence split, overview mosaic, process roadmap, comparison, or matrix from the template's verified body variants. Thin and placeholder pages remain explicit review items; generic filler should not be used to make a page look full.

For Chinese硕博士毕业答辩, pass --scenario-profile thesis_defense to activate scenario_variant cn_degree_defense_v4. The intake then records thesis chapter hierarchy, duration-based page bands, an A-D source-figure index, high-resolution figure review flags, and the four stages source planning -> editable master -> editable sample -> full editable deck. Image2 is off by default; use the explicit --visual-exploration flag only for upstream style exploration.

Single-Paper Literature-Report Flow Selection

Before deriving the outline for single_paper_report, read references/literature-report-flow-selection.md.

If the user provides a page outline, speaking script, learning notes, or slide-by-slide plan, that structure is the primary story contract. Preserve the user's page order and speaking logic, then use the paper/SI to verify claims and source figures. Do not replace the user's structure with an automatic long-form literature-report flow unless the user explicitly asks for a rebuild.

If no outline or script is provided, choose or present two flow options:

• paper_ppt_concise: concise, figure-first 6-10 slide paper report inspired by xiao634zhang/paper-ppt-skill.
• literature_report_deep_dive: 20+ slide deep literature-report planning flow inspired by fangyuanopus/literature-report-ppt-builder, with figure_source_manifest, deck_order_map, and page_briefs before visual execution.

Both options still use EasySlides' editable SVG/shape-IR -> DrawingML backend as the production path.

Scenario Profiles and Rule Layers

Before the confirmation step for academic decks, load the scenario profile catalog from references/scenario_profiles.json.

python scripts/scenario_profiles.py --list --json
python scripts/scenario_profiles.py --profile <profile_id> --json

Seed academic profiles are single_paper_report, multi_paper_review, thesis_defense, proposal_or_fund, lab_progress, workshop_training, and conference_talk. Pick the nearest profile from the source material, audience, and occasion, then state it as a recommendation during confirmations. If the deck is an academic scenario outside these seeds, keep the nearest profile as the rule base and record a scenario_variant rather than squeezing the deck into defense or literature-report wording.

Chinese NSFC Proposal Defense

When the user asks for a Chinese National Natural Science Foundation proposal defense and selects nsfc_defense, activate scenario_id: nsfc_grant_cn. This is a user-facing intent, not a request for the user to learn plan fields: the agent records the scenario contract internally and keeps the conversation in natural language.

For a full deck, organize the argument as: (1) rationale and scientific question, (2) three research contents and their technical route, and (3) innovation, feasibility, and implementation plan. Every page must have one declared narrative responsibility, and the template binds that responsibility to a reviewed content variant. A table-of-contents page is optional; never add one merely because the template has a TOC shell.

Ask with explicit choices before generation only when the request leaves a result-affecting decision unresolved, such as whether this is a project application or a progress defense, whether the user wants a full or concise deck, or whether supplied material may be supplemented by external research. Do not ask the user for grant_role, body_variant_id, coordinates, or other implementation details.

Apply rule layers in this order:

• hard_rules: always enforce; source faithfulness, traceability, citations, text fit, template geometry integrity, and PPTX deliverability.
• required_rules: enforce for the selected scenario unless the user changes the scenario.
• recommended_rules: default guidance such as action titles or ghost-deck checks; apply when useful, but let templates and content needs override.
• relaxable_rules: soft defaults that can yield to venue, template, or teaching/activity format.

Templates control visual containers; scenario profiles control content organization. Template colors, title treatment, page chrome, icon style, and layout density may override profile recommendations only when allowed by template_may_override; they must not override protected items in template_must_not_override.

Deck Plan Contract

After confirmations and before spec_lock.md, write deck_plan.json as the page-level academic story contract:

python scripts/deck_plan_contract.py <project_path>/deck_plan.json --json

Each slide entry must include page, role, action_title, claim, evidence_sources, layout_id, rhythm, and speaker_note. Use action_title for the page's conclusion sentence, claim for the supported statement, and evidence_sources for source-map references such as paper pages, figure/table ids, extracted images, datasets, or user-provided assets. The validated deck plan feeds design_spec.md, spec_lock.md page_rhythm, and later QA gates.

After the deck plan passes, freeze the execution handoff as deck_execution_lock.json:

python scripts/deck_execution_lock.py <project_path>/deck_plan.json --write <project_path>/deck_execution_lock.json --json

The execution lock records the per-page layout_id, rhythm, evidence refs, body variant selection, declared slots, template palette, and required gates. Executor must re-read deck_execution_lock.json together with spec_lock.md before each page. If either file disagrees with the current deck plan, validate with scripts/deck_execution_lock.py --validate and fix the plan before SVG authoring.

When the selected template has body_variants.json, every executable content slide must use a verified body variant in layout_id, provide content_shape, and fill slot_payload with exactly the slots declared by that variant. The scripts/deck_plan_contract.py report includes body_variant_status from scripts/body_variant_adapter.py; body_variant_contract, template_tokens, text_capacity, svg_quality_checker, preview_render, pptx_roundtrip, and visual_measure_gate are blocking gates before SVG authoring. Do not let Executor bypass this by writing arbitrary SVG into a template body canvas.

For managed template packages, compile the canonical sources before rendering:

python scripts/easyslides.py template-compile templates/layouts/<template_id> --write --json
python scripts/easyslides.py slide-compile <project_path>/deck_plan.json --template <template_id> --out <project_path>/slide_ir.json --svg-out <project_path>/svg --pptx-out <project_path>/exports/<deck>.pptx --json

Treat template_package.json, layouts.json, body_variants.json, component_catalog.json, and qa_policy.json as owned sources. Treat compiled/template_ir.json, compiled/template.lock.json, flattened contracts, status files, and registries as generated projections. A composable body variant must bind each required component-local slot to a declared variant slot and resolve every instance to a named region or explicit placement. If several variants score equally and the user intent is unclear, block execution and ask the user to choose; do not silently pick the first variant. A template may expose open_component_composition for an explicitly requested arrangement, but that route accepts only registered local components inside the declared body canvas and never permits global or cross-template assets.

When the selected template exposes a LOGO slot, resolve it from user-provided assets and source-material image folders before SVG generation. Prefer an actual institutional, laboratory, project, or paper-source logo. If no suitable logo is available, keep the template's built-in degree-cap icon fallback; when a real logo is inserted, replace the full LOGO group or hide the fallback drawing.