Ppt Master

PPT Master Skill

AI-driven multi-format SVG content generation system. Converts source documents into high-quality SVG pages through multi-role collaboration and exports to PPTX.

Core Pipeline: Source Document → Create Project → [Template] → Strategist → [Image_Generator] → Executor Live Preview → Quality Check → Post-processing → Export

[!CAUTION]

🚨 Global Execution Discipline (MANDATORY)

This workflow is a strict serial pipeline. The following rules have the highest priority — violating any one of them constitutes execution failure:

1. SERIAL EXECUTION — Steps MUST be executed in order; the output of each step is the input for the next. Non-BLOCKING adjacent steps may proceed continuously once prerequisites are met, without waiting for the user to say "continue"
2. BLOCKING = HARD STOP — Steps marked ⛔ BLOCKING require a full stop; the AI MUST wait for an explicit user response before proceeding and MUST NOT make any decisions on behalf of the user
3. NO CROSS-PHASE BUNDLING — Cross-phase bundling is FORBIDDEN. (Note: the Eight Confirmations in Step 4 are ⛔ BLOCKING — the AI MUST present recommendations and wait for explicit user confirmation before proceeding. Once the user confirms, all subsequent non-BLOCKING steps — design spec output, SVG generation, speaker notes, and post-processing — may proceed automatically without further user confirmation)
4. GATE BEFORE ENTRY — Each Step has prerequisites (🚧 GATE) listed at the top; these MUST be verified before starting that Step
5. NO SPECULATIVE EXECUTION — "Pre-preparing" content for subsequent Steps is FORBIDDEN (e.g., writing SVG code during the Strategist phase)
6. NO SUB-AGENT SVG GENERATION — Executor Step 6 SVG generation is context-dependent and MUST be completed by the current main agent end-to-end. Delegating page SVG generation to sub-agents is FORBIDDEN
7. SEQUENTIAL PAGE GENERATION ONLY — In Executor Step 6, after the global design context is confirmed, SVG pages MUST be generated sequentially page by page in one continuous pass. Grouped page batches (for example, 5 pages at a time) are FORBIDDEN
8. SPEC_LOCK RE-READ PER PAGE — Before generating each SVG page, Executor MUST read_file <project_path>/spec_lock.md. All colors / fonts / icons / images MUST come from this file — no values from memory or invented on the fly. Executor MUST also look up the current page's page_rhythm (anchor / dense / breathing), page_layouts (which template SVG to inherit, if any), and page_charts (which chart template to adapt, if any). Empty / absent entries are intentional Strategist signals — see executor-base.md §2.1. This rule exists to resist context-compression drift on long decks and to break the uniform "every page is a card grid" default
9. SVG MUST BE HAND-WRITTEN, NOT SCRIPT-GENERATED — Every SVG page is written by the main agent directly, one page at a time (see rules 6 and 7). Writing or running a Python / Node / shell script that produces the SVG files in batch — looping over pages, templating from data, or emitting them via a generator — is FORBIDDEN, including under "save tokens", "quick draft", or "user is in a hurry" pretexts. The script-generation path was tried on a feature branch and abandoned: cross-page visual consistency depends on per-page authoring with full upstream context, which a generator script cannot reproduce

[!IMPORTANT]

🌐 Language & Communication Rule

• Response language: match the user's input and source materials. Explicit user override (e.g., "请用英文回答") takes precedence.
• Template format: design_spec.md MUST follow its original English template structure (section headings, field names) regardless of conversation language. Content values may be in the user's language.

[!IMPORTANT]

🔌 Compatibility With Generic Coding Skills

• ppt-master is a repository-specific workflow, not a general application scaffold
• Do NOT create .worktrees/, tests/, branch workflows, or generic engineering structure by default
• On conflict with a generic coding skill, follow this skill unless the user explicitly says otherwise

Main Pipeline Scripts

For complete tool documentation, see ${SKILL_DIR}/scripts/README.md.

Windows note: if a python3 ... command fails (common on python.org installs, which provide python.exe but not python3.exe), rerun the same command with python instead.

Template Index

Standalone Workflows

Workflow

Step 1: Source Content Processing

🚧 GATE: User has provided source material (PDF / DOCX / EPUB / URL / Markdown file / text description / conversation content — any form is acceptable).

No source content? When the user supplies only a topic name or requirements without any file or substantive description, run the topic-research workflow first, then return here with its products as input.

When the user provides non-Markdown content, convert immediately:

Office vector assets (EMF/WMF) from DOCX/PPTX sources: doc_to_md.py / ppt_to_md.py extract embedded Office vector images (.emf/.wmf) alongside bitmap images. After import-sources, these land in images/ together with image_manifest.json and are first-class assets in §VIII Image Resource List.

Do NOT convert EMF/WMF to PNG. The PPT Master pipeline preserves them as external references (finalize_svg.py skips them) and svg_to_pptx.py embeds them as PPTX-native media via image/x-emf / image/x-wmf MIME — PowerPoint renders them at full vector fidelity. Converting via LibreOffice/Inkscape introduces CJK font substitution drift and rasterization loss; the original EMF/WMF is always higher fidelity than the converted PNG.

Browser-based live preview cannot render EMF (will show blank) — this is expected; the PPTX output is the source of truth.

✅ Checkpoint — Confirm source content is ready, proceed to Step 2.

Step 2: Project Initialization

🚧 GATE: Step 1 complete; source content is ready (Markdown file, user-provided text, or requirements described in conversation are all valid).

python3 ${SKILL_DIR}/scripts/project_manager.py init <project_name> --format <format>

Format options: ppt169 (default), ppt43, xhs, story, etc. For the full format list, see references/canvas-formats.md.

Import source content (choose based on the situation):

⚠️ MUST use --move (not copy): all source files — Step 1's generated Markdown, original PDFs / MDs / images — go into sources/ via import-sources --move. After execution they no longer exist at the original location. Intermediate artifacts (e.g., _files/) are handled automatically.

✅ Checkpoint — Confirm project structure created successfully, sources/ contains all source files, converted materials are ready. Proceed to Step 3.

Step 3: Template Option

🚧 GATE: Step 2 complete; project directory structure is ready.

Default — free design. Proceed directly to Step 4. Do NOT query any *_index.json unless triggered. Do NOT ask the user. Do NOT proactively suggest, hint at, or fuzzy-match any template based on content, slug-like words, or vague style descriptions.

Template flow triggers ONLY on explicit directory paths supplied by the user in their initial message. The trigger rule is mechanical, not interpretive:

There is no slug matching, no name lookup, no fuzzy resolution. A name without a path does not trigger — the user must give a path the AI can cd into.

Style descriptions ("麦肯锡风格" / "Keynote 风" / "极简风" / etc.) never trigger Step 3. They flow into Strategist's Eight Confirmations as a style brief (color / typography / tone in confirmations e–g).

Bare names ("academic_defense", "招商银行", "anthropic") do NOT trigger Step 3 even if a matching directory exists in the library. The user must give a path. AI must not "helpfully" resolve a name to a path.

"What templates exist?" is out-of-band Q&A — answer by listing entries from brands_index.json / layouts_index.json / decks_index.json together with their paths. Listing alone does not advance the pipeline; the user must send a path back to trigger Step 3.

To create a new layout or deck, read workflows/create-template.md. To create a new brand, read workflows/create-brand.md.

Three template kinds

The architecture has three independent reference bundles. Full schema in docs/zh/templates-architecture.md. Summary:

Segment ownership (governs fusion override priority):

Single-path dispatch

TEMPLATE_DIR=<user-supplied path>
cp -r ${TEMPLATE_DIR}/* <project_path>/templates/

The single-line copy suffices for all three kinds — the spec's kind field tells Strategist how to read it; downstream code doesn't distinguish.

Multi-path fusion

When the user gives two or more paths of different kinds, Step 3 fuses them into a single <project>/templates/design_spec.md. Default granularity is segment-level integer replacement — entire identity / structure / middle segments are taken from the highest-priority source for that segment, no implicit field-level mixing.

Override priority by segment:

Field-level micro-adjustment (e.g. "use anthropic brand but primary changed to #FF0000") is not part of Step 3 fusion — it flows into Strategist Eight Confirmations e–g as a normal user request.

Same-kind multiple paths — conflict resolution

When the user gives two paths of the same kind (e.g. brands/anthropic + brands/google), Step 3 surfaces a conflict prompt before fusing — like resolving a git merge conflict:

AI: 你给了两个 brand，检测到段级冲突：
    - Color Scheme（Anthropic 橙红 vs Google 多色）
    - Typography（Styrene/AnthropicSans vs GoogleSans/Roboto）
    - Logo（Anthropic 标 vs Google 标）
    - Voice & Tone（restrained vs friendly）
    - Icon Style（stroke vs filled）

    要 (a) 全部按 Anthropic / (b) 全部按 Google / (c) 逐段挑？

Rules:

• Default: no implicit ordering — every cross-source segment difference is reported as a conflict
• Only when the user picks (c) does AI walk through each segment one by one
• Field-level conflicts are out of scope — segment-level only
• Three or more same-kind paths are not supported — ask the user to converge to at most two

Fused spec provenance

When fusion happens (any multi-path case), the resulting <project>/templates/design_spec.md carries a provenance block immediately under its H1:

> **Fused from:**
> - deck: `templates/decks/招商银行/` （base）
> - brand: `templates/brands/anthropic/` （identity override）
> - layout: `templates/layouts/academic_defense/` （structure override）
> - conflicts resolved: Color Scheme from anthropic（user picked a）

Single-path Step 3 does not add provenance (the source is self-evident from the copied files).

✅ Checkpoint — Default path proceeds to Step 4 without user interaction. If the user supplied one or more explicit template paths, those have been dispatched (or fused) into <project_path>/templates/ before advancing.

Step 4: Strategist Phase (MANDATORY — cannot be skipped)

🚧 GATE: Step 3 complete; default free-design path taken, or (if triggered) template files copied into the project.

First, read the role definition:

Read references/strategist.md

⚠️ Mandatory gate: before writing design_spec.md, Strategist MUST read_file templates/design_spec_reference.md and follow its full I–XI section structure. See strategist.md Section 1.

Eight Confirmations (full template: templates/design_spec_reference.md):

⛔ BLOCKING: present the Eight Confirmations as a single bundled recommendation set and wait for explicit user confirmation or modification before outputting Design Specification & Content Outline. This is the single core confirmation point — once confirmed, all subsequent steps proceed automatically.

1. Canvas format
2. Page count range
3. Target audience
4. Style objective
5. Color scheme
6. Icon usage approach
7. Typography plan, including formula rendering policy
8. Image usage approach

Mandatory — split-mode note (not a ninth confirmation): after listing the eight confirmation details, you MUST append exactly one short line (rendered in the user's language, prefixed with 💡) about generation mode. Pick the variant by qualitative read of Phase A signals — recommended page count, source-material bulk, whether topic-research ran with substantial web-fetch accumulation:

This line is required output every run — the user must always see the mode choice exists. Whether to act on it is the user's call.

Formula rendering policy lives inside item 7 (Typography plan):

After the Eight Confirmations are approved and before outputting design_spec.md / spec_lock.md, if the confirmed formula policy is mixed or render-all and the content contains formula-worthy expressions, Strategist MUST:

1. Identify explicit LaTeX and any source expressions that should be faithfully structured as formulas.
2. Write <project_path>/images/formula_manifest.json with only the formulas selected for rendering.
3. Run:

   python3 ${SKILL_DIR}/scripts/latex_render.py <project_path>
   

4. Include the rendered formula PNGs as Acquire Via: formula, Status: Rendered, Type: Latex Formula rows in design_spec.md §VIII Image Resource List; also list them in spec_lock.md images with | no-crop.

The formula renderer uses a provider fallback chain by default: codecogs,quicklatex,mathpad,wikimedia. The first three are color-aware; Wikimedia is an availability fallback. Formula PNGs are transparent by default: manifest background is the temporary render matte and transparency-removal reference, not a retained final background unless transparent: false is set for that item. Do not scan spec_lock.md for $...$ or $$...$$. Dollar-delimited math in source material is only a signal for Strategist; the renderer consumes the explicit manifest.

If the user provided images or formula PNGs were rendered, run analysis before outputting the design spec:

python3 ${SKILL_DIR}/scripts/analyze_images.py <project_path>/images

⚠️ Image handling: NEVER directly read / open / view image files (.jpg, .png, etc.). All image info comes from analyze_images.py output or the Design Spec's Image Resource List.

Output:

• <project_path>/design_spec.md — human-readable design narrative
• <project_path>/spec_lock.md — machine-readable execution contract (skeleton: templates/spec_lock_reference.md); Executor re-reads before every page

✅ Checkpoint — Phase deliverables complete, auto-proceed to next step:

## ✅ Strategist Phase Complete
- [x] Eight Confirmations completed (user confirmed)
- [x] Split-mode note appended below the eight items (heavy or normal variant)
- [x] Design Specification & Content Outline generated
- [x] Execution lock (spec_lock.md) generated
- [ ] **Next**: Auto-proceed to [Image_Generator / Executor] phase

Step 5: Image Acquisition Phase (Conditional)

🚧 GATE: Step 4 complete; Design Specification & Content Outline generated and user confirmed. Any formula rows already have Acquire Via: formula and Status: Rendered.

Trigger: At least one row in the resource list has Acquire Via: ai and/or Acquire Via: web. If every row is user, formula, or placeholder, skip to Step 6.

Always load the common framework:

Read references/image-base.md

Then lazy-load the path-specific reference for each row that actually needs it:

A deck with only ai rows never loads image-searcher.md; a deck with only web rows never loads image-generator.md. A mixed deck loads both, processes each row through its own path, and writes both image_prompts.json and image_sources.json.

⚠️ In-pipeline ai path MUST use manifest mode — even when only 1 ai row exists. Write images/image_prompts.json first, then run image_gen.py --manifest, then image_gen.py --render-md to produce the image_prompts.md sidecar. The positional form (image_gen.py "prompt" ...) is reserved for out-of-pipeline one-off testing / single-image fixups — it skips manifest + sidecar, leaving no audit trail.

Workflow:

1. Extract all rows with Status: Pending and Acquire Via ∈ {ai, web} from the design spec
2. Generate prompts (ai rows) and/or run search (web rows) per image-base.md §2 dispatch table
3. Verify every row reaches a terminal status: Generated (ai success), Sourced (web success), or Needs-Manual

✅ Checkpoint — Confirm acquisition attempted for every row:

## ✅ Image Acquisition Phase Complete
- [x] image_prompts.json created (when any ai rows processed)
- [x] image_prompts.md sidecar rendered (when any ai rows processed)
- [x] image_sources.json created (when any web rows processed)
- [x] Each row: status is `Generated` / `Sourced` / `Needs-Manual` (no `Pending` remaining)

Default — auto-proceed to Step 6. Only when the user's Step 4 response explicitly opted into split mode (in reply to the optional hint), output the Phase A hand-off below and stop this conversation:

## ✅ Phase A Complete
- [x] Spec: `design_spec.md`, `spec_lock.md`
- [x] Resources: `sources/`, `images/`, `templates/`
- [ ] **Next**: open a fresh chat window and input `继续生成 projects/<project_name>` to enter Phase B via the [`resume-execute`](workflows/resume-execute.md) workflow.

On acquisition failure, do NOT halt — follow the Failure Handling rule in image-base.md §5: retry once, then mark the row Needs-Manual, report to user, and continue to the checkpoint above.

Step 6: Executor Phase

🚧 GATE: Step 4 (and Step 5 if triggered) complete; all prerequisite deliverables are ready.

Read the role definition based on the selected style:

Read references/executor-base.md          # REQUIRED: common guidelines
Read references/shared-standards.md       # REQUIRED: SVG/PPT technical constraints
Read references/executor-general.md       # General flexible style
Read references/executor-consultant.md    # Consulting style
Read references/executor-consultant-top.md # Top consulting style (MBB level)

Only read executor-base + shared-standards + one style file.

Design Parameter Confirmation (Mandatory): before the first SVG, output key design parameters from the spec (canvas dimensions, color scheme, font plan, body font size). See executor-base.md §2.

Live Preview Auto-Startup (Mandatory): before the first SVG, automatically start the browser editor in live mode and keep it running continuously through Executor + Step 7 export:

python3 ${SKILL_DIR}/scripts/svg_editor/server.py <project_path> --live

• Start it immediately when Executor begins; svg_output/ may be empty. Editor opens at http://localhost:5050; port conflict → --port <other> and report the actual URL.
• Run it as a long-running side process/session; do not wait for it to exit before generating SVG pages. Do not wait for user confirmation after startup.
• Service must keep running until one of: (a) the user clicks Exit preview in the browser, or (b) the user explicitly asks in chat to stop it. Generation continues even if the user closes the editor.
• Do NOT read or apply submitted annotations during generation. Users may annotate at any time, but Executor proceeds without touching them. The window to apply annotations opens only after Step 7 completes — see workflows/live-preview.md.
• The editor also supports staged direct edits (text content + SVG element attributes previewed immediately, then written to svg_output/ only when the user clicks Apply changes; Ctrl+Z / Undo drops staged edits) alongside annotation; re-export stays chat-driven. Full scope and editor details: see workflows/live-preview.md Notes.

Pre-generation Batch Read (Mandatory): before the first SVG, batch-read every distinct layout SVG referenced in spec_lock.page_layouts and every distinct chart SVG referenced in spec_lock.page_charts (plus any §VII backup charts). One read per file, up front — do not re-read these during page generation. See executor-base.md §1.0.

Per-page spec_lock re-read (Mandatory): before each SVG page, read_file <project_path>/spec_lock.md and use only its colors / fonts / icons / images, plus the per-page page_rhythm / page_layouts / page_charts lookups (resolves to template SVGs already loaded in the batch read above). Resists context-compression drift on long decks. See executor-base.md §2.1.

⚠️ Main-agent only: SVG generation MUST stay in the current main agent — page design depends on full upstream context. Do NOT delegate to sub-agents. ⚠️ Generation rhythm: generate pages sequentially, one at a time, in the same continuous context. Do NOT batch (e.g., 5 per group).

Visual Construction Phase: generate SVG pages sequentially, one at a time, in one continuous pass → <project_path>/svg_output/

Quality Check Gate (Mandatory) — after all SVGs, BEFORE annotation handling and speaker notes:

python3 ${SKILL_DIR}/scripts/svg_quality_checker.py <project_path>

• Any error (banned SVG features, viewBox mismatch, spec_lock drift, etc.) MUST be fixed before proceeding — return to Visual Construction, regenerate that page, re-run check.
• warning entries (low-res image, non-PPT-safe font tail, etc.): fix when straightforward, otherwise acknowledge and release.
• Run against svg_output/ (not after finalize_svg.py — finalize rewrites SVG and masks violations).

Logic Construction Phase: generate speaker notes → <project_path>/notes/total.md

✅ Checkpoint — Confirm all SVGs and notes are fully generated and quality-checked. Proceed directly to Step 7 post-processing:

## ✅ Executor Phase Complete
- [x] Live preview started and kept available at the reported URL
- [x] All SVGs generated to svg_output/
- [x] svg_quality_checker.py passed (0 errors)
- [x] Speaker notes generated at notes/total.md

Chart pages? If this deck contains data charts (bar / line / pie / radar / etc.), run the standalone verify-charts workflow before Step 7 to calibrate coordinates. AI models routinely introduce 10–50 px errors when mapping data to pixel positions; verify-charts eliminates that class of error. Skip if no chart pages.

Visual self-check (opt-in)? If the user explicitly asked for a per-page visual re-pass on the SVGs ("跑一下视觉自检 / 视觉回看", "visual review", "check pages visually", etc.), run the standalone visual-review workflow before Step 7. Do NOT run it by default and do NOT recommend it based on inferred model capability or deck size — trigger is user request only.

Step 7: Post-processing & Export

🚧 GATE: Step 6 complete; all SVGs generated to svg_output/; speaker notes notes/total.md generated.

🚧 Image readiness GATE (when Step 5 left ai rows in Needs-Manual): every expected file must exist at project/images/<filename> before running 7.1.

If files are missing: PAUSE, list the missing filenames, point the user to images/image_prompts.md (each ### Image N: block is paste-ready for ChatGPT / Gemini / Midjourney; auto-generated from image_prompts.json) and the required placement project/images/<filename>. Resume Step 7.1 only after all expected files are in place. finalize_svg.py and svg_to_pptx.py do not detect missing files at this layer — proceeding with gaps produces a deck with broken image references.

⚠️ Run the three sub-steps one at a time — each must complete successfully before the next. ❌ NEVER combine them into a single code block or shell invocation.

Canonical three-command pipeline (mirrors references/shared-standards.md §5):

Step 7.1 — Split speaker notes:

python3 ${SKILL_DIR}/scripts/total_md_split.py <project_path>

Step 7.2 — SVG post-processing (icon embedding / image crop & embed / text flattening / rounded rect to path):

python3 ${SKILL_DIR}/scripts/finalize_svg.py <project_path>

Step 7.3 — Export PPTX (embeds speaker notes by default):

python3 ${SKILL_DIR}/scripts/svg_to_pptx.py <project_path>
# Output (default-flow mode):
#   exports/<project_name>_<timestamp>.pptx           ← native pptx (canonical output, reads svg_output/)
#   backup/<timestamp>/svg_output/                    ← Executor SVG source backup (always written)
#
# Add --svg-snapshot to additionally emit the SVG-image preview pptx alongside the native pptx:
#   exports/<project_name>_<timestamp>_svg.pptx      ← SVG preview pptx (reads svg_final/)

The native pptx consumes svg_output/ directly so the converter can preserve high-fidelity primitives (icon <use> placeholders, image preserveAspectRatio → srcRect, rounded rect rx/ry → prstGeom roundRect). The svg_output/ snapshot in backup/<timestamp>/ is always written so the project can be re-exported from frozen SVG sources without re-running the LLM. The SVG-rendered preview pptx is opt-in via --svg-snapshot — live preview already provides the SVG visual reference, so it's only needed when you want a self-contained file to share. Pass -s output or -s final to force a single source if you need it.

Paragraph editability vs line fidelity — by default every dy-stacked line is its own PowerPoint text frame, preserving exact SVG layout. Add --merge-paragraphs only when the user explicitly asks for an editable / wrap-friendly export (e.g. "I want to edit the abstract as one block", "make text boxes resizable / reflow"): mergeable paragraph blocks collapse into one editable text frame with multiple <a:p>, at the cost of PowerPoint re-wrapping inside each box. Default off keeps pixel-fidelity; turn it on per the user's request, not on your own judgement.

Optional animation flags (the defaults already enable rich entrance animations — adjust only when the user asks for something different):

• -t <effect> — page transition. Default fade. Options: fade / push / wipe / split / strips / cover / random / none.
• -a <effect> — per-element entrance animation. Default auto (map effect from group id: chart→wipe, card-/step-/pillar-→fly, title/takeaway→fade; image-like ids hero / figure- / image / img- / kpi cycle a richer pool — zoom / dissolve / circle / box / diamond / wheel — so multiple images vary across the deck). Pass none to disable, a specific effect like fade, or mixed for the legacy 16-effect cycle. Requires top-level <g id="..."> groups (already required by Executor).
• --animation-trigger {on-click,with-previous,after-previous} — Start mode (matches PowerPoint's animation-pane Start dropdown). Default after-previous (click-free cascade; pace via --animation-stagger). Use on-click for presenter-paced reveals, or with-previous for all-at-once.
• --animation-config <path> — optional object-level sidecar. Default: <project_path>/animations.json when present.
• --auto-advance <seconds> — kiosk-style auto-play.

Optional custom animations (only when the user asks to tune animation order/effects/timing for specific objects):

Run the standalone customize-animations workflow. Default export already has global entrance animation; do not create animations.json unless object-level customization was requested.

Optional recorded narration (only when the user asks for narrated/video export):

Run the standalone generate-audio workflow. The AI picks a narration backend (edge by default, or a configured cloud provider such as ElevenLabs / MiniMax / Qwen / CosyVoice for high-quality or cloned voices), asks the user once (backend + voice + rate/settings + embed-or-not, all with recommended values), then executes notes_to_audio.py and (if chosen) re-exports the PPTX with --recorded-narration audio.

Do NOT call notes_to_audio.py directly without going through the workflow — --voice / --voice-id is required and the workflow produces the locale/provider-aware recommendation that makes the choice meaningful.

Full effect list, anchor logic, and limits: references/animations.md.

❌ NEVER substitute cp for finalize_svg.py — finalize performs multiple critical processing steps ❌ NEVER force -s output for the legacy/preview pptx (PowerPoint's internal SVG parser drops icons and rounded corners). The default auto-split already gives native the high-fidelity source it needs without touching legacy. ❌ NEVER use --only (it suppresses one of the two output files)

Post-export annotation window: the preview service from Step 6 typically remains running after export. If the user submitted annotations in the browser (during Executor or after export) and now asks to apply them — they may quote the browser prompt (Changes saved to svg_output... / 修改已保存到 svg_output...), say "apply my annotations" / "应用注解" / equivalent — run live-preview Step 2 to apply and re-export. Annotations submitted during generation are also handled here, not earlier.

Direct edits in the browser: the user may also stage text / SVG attribute edits in the preview. These land in svg_output/ only after the user clicks Apply changes. If they ask to "re-export" / "重新导出" after applying such edits, just re-run Step 7.2–7.3 (finalize + export); no annotation-application step is needed unless they also saved AI-needed annotations.

Preview not running? Any time the user mentions "live preview", "preview", "看效果", or wants to select/click a slide element and the service is not running, run live-preview Step 1 to start it. If the service is already running, just point them at the URL — do not restart.

Role Switching Protocol

Before switching roles, MUST first read the corresponding reference file. Output marker:

## [Role Switch: <Role Name>]
📖 Reading role definition: references/<filename>.md
📋 Current task: <brief description>

Reference Resources

Notes

• Local preview: python3 -m http.server -d <project_path>/svg_final 8000
• Troubleshooting: on generation issues (layout overflow, export errors, blank

…(truncated)