| name | ppt-image-first |
| description | Build presentation plans for PPT / slides / decks through a conversation-first workflow, then propose multiple visual directions with preview images before writing deck specs. Use when the user asks to create a PPT, presentation, deck, 答辩稿, 路演 deck, 产品介绍 PPT, 汇报 PPT, or when the user has only a topic or rough materials and needs help clarifying structure, style, and page planning before generation. |
PPT Image First
Use this skill to turn a vague PPT request into a generation-ready, image-first deck plan.
I/O Contract
- Input: a PPT topic, rough goal, existing notes/materials, or complete report-like narrative.
- Optional anchors: audience, page count or duration, school/company/lab/course/brand identity, use occasion, source files, and style constraints.
- Output: a confirmed content basis, visual style previews, deck specs, generated page visuals, review surface, and final PPT after approval.
- Confirmation gates:
需求确认, 风格确认, 生成前确认, and final review approval.
- Default ratio: generate preview images and final page visuals in
16:9 unless the user explicitly requests another ratio.
Workflow Phases
Follow these phases in order and do not skip the confirmation gates:
- Intake: clarify lightly, identify anchors, output a short baseline judgment, then get
需求确认.
- Content basis: create or extract
content_report.md unless the user already supplied complete report-like content; show a synthesized judgment, not the raw file by default.
- Style preview: align style boundaries, propose directions, generate
首页 / 目录页 / 正文页 previews for each direction, support refinement, then get 风格确认.
- Planning lock: run
风格反演确认, then write design_spec.md, slide_blueprint.md, and spec_lock.md in that order; summarize and get 生成前确认.
- Generation review: generate page visuals, optionally run candidate selection, open the review HTML, retouch as needed, and export the final PPT only after approval.
Read references/workflow.md for stage-by-stage execution details.
Progressive Loading
- Intake and baseline: read
references/conversation_framework.md only when you need the conversation-first intake pattern.
- Style proposals: read
references/style-system.md when producing proposal cards or interpreting V1-V8 internally.
- Preview/review UI: read
references/preview-flow.md before producing style previews, candidate-picker pages, or review pages.
- Content basis: read
templates/content_report_reference.md before writing content_report.md.
- Planning files: read
templates/design_spec_reference.md, templates/slide_blueprint_reference.md, and templates/spec_lock_reference.md immediately before writing those files.
Working Principles
- Treat the user like a client and yourself like the proposing design agent.
- Do not force the user to configure design parameters manually; translate their natural-language intent into design decisions.
- Ask for real-world identity anchors when relevant, but treat them as grounding context rather than style questions.
- Keep intake lightweight; do not turn content basis or style alignment into a long questionnaire.
- Use style vectors internally, but present front-stage options as concise proposal cards plus image previews.
- Use the content basis for previews and planning rather than generic placeholder structures.
- Mark inferred claims carefully; do not present unsupported precise data, experiment results, citations, or institutional conclusions as user-provided facts.
- Show summaries, proposal cards, previews, and review surfaces by default; show raw planning files only when useful or requested.
Hard Rules
- Preview-first: final style confirmation should be based on generated
首页 / 目录页 / 正文页 previews, not text-only choices.
- Shells are mandatory: use
assets/preview_shell/index.html, assets/candidate_picker_shell/index.html, and assets/review_shell/index.html as the base UIs for their stages.
- Open the local preview, candidate-picker, review, and final PPT files immediately when those stages produce them; if opening fails, state the blocker and provide the path.
- Do not write
slide_blueprint.md before style direction and 风格反演确认 are complete.
- Keep generation image-first when the confirmed direction depends on generated page visuals; do not silently fall back to shape-by-shape PPT construction or hand-drawn code substitutes.
- User-requested edits such as adding, replacing, correcting, or supplementing visible text are still image-generation/editing tasks by default. Do not use PIL, Pillow, canvas, SVG, HTML screenshots, PPT native text boxes, or deterministic overlays to patch final page visuals unless the user explicitly asks for a code/PPT-overlay workaround.
- Treat generated page visuals as complete by default; post-generation overlays default to zero unless they are traceable to approved blueprint fields.
- Keep slide identifiers, candidate codes, filenames, and generation batch labels outside the image-generation prompt body. They may appear in planning files, filenames, mapping tables, review UI, and chat instructions, but the prompt sent to the image model should contain only audience-facing content and visual direction.
- Review payloads should carry lightweight coordinate markup, not base64 preview images. When a user pastes review JSON with markup, first render the marked review images locally with
scripts/render_review_markup.py, then use the marked images plus separate text comments as the retouch/regeneration reference.
- Before full generation, ask whether the user wants one final image per slide or multiple final candidates per slide.
- Do not export the final PPT immediately after first-pass generation; export only after the reviewed pages are approved.
Output Hierarchy
Use these layers consistently:
content_report.md = upstream content basis when the user did not provide complete report-like material.
design_spec.md = global deck rationale and confirmed visual system.
slide_blueprint.md = page-by-page intent and content plan.
spec_lock.md = execution constraints and final generation guardrails.
content_report.md supports the 3 core planning files; it does not replace them. spec_lock.md is always the last core planning file.