Skip to content

SKILL.md

Latest commit

 

History

History
106 lines (84 loc) · 6.44 KB

SKILL.md

File metadata and controls

106 lines (84 loc) · 6.44 KB
Raw
name ppt-image-share-builder
description Build image2-first classroom PPT page images from a topic, source files, reference PPT style, and iterative user feedback. Use when Codex needs to create image2-ready per-page prompts, generate polished PPT page images, QA a contact sheet, insert the final images into a PPTX wrapper, and write a timed presentation script for a course report or lecture-style PPT.

PPT Image Share Builder

Use this skill for image2-first PPT image work:

source files + reference PPT style
  -> sourced page outline
  -> image2-ready per-page prompts
  -> generated 16:9 PPT page images
  -> contact-sheet QA
  -> targeted fixes only if QA fails or user asks
  -> full-bleed PPTX wrapper
  -> timed speaking script

Keep the workflow source-backed, style-aware, and human-reviewed. Helper scripts are for post-generation steps; they do not replace image2.

Core Rules

  • Treat generated PPT page images as the visual source of truth. The PPTX is only a delivery wrapper.
  • Optimize for a non-expert user. Do not require the user to know prompt engineering terms; translate vague feedback into concrete content, style, and cleanup actions.
  • Before writing prompts or generating images, run bounded grill-me style intake. If a real grill-me skill/tool is available, use it; otherwise use references/intake-grillme.md.
  • Do not begin deck planning or image generation until essential decision branches are resolved or explicit assumptions are recorded.
  • Extract and verify source facts before writing prompts. Use official or primary sources for recent rules, cases, data, or dates.
  • Match the user's reference PPT style: layout rhythm, title system, color palette, page markers, diagram grammar, and information density.
  • Default to one-pass delivery: generate the full useful deck in one run unless the user explicitly asks for staged confirmation or the request is too ambiguous to proceed safely.
  • Use internal gates instead of user-visible rounds: content depth gate, style lock gate, and cleanup QA gate.
  • Never let the first delivered version be merely decorative. Each non-cover/non-closing page needs a claim, 2-4 substantive points, and a meaningful visual structure.
  • Create a style lock before generating the full image set; every page prompt must repeat that style lock.
  • In final QA, remove meaningless elements: stray X/check marks, random labels, fake legends, unrelated icons, decorative charts with no data role, duplicate badges, and any unrequested Q&A text.
  • Keep generated Chinese text short. Recommend adding exact long text manually in PPT if precision matters.
  • For dense Chinese or source-critical text, use image2 for the visual page or background and compose exact text locally, then treat the exported full-page image as the final slide.
  • Do not generate government seals, company logos, QR codes, or watermarks unless the user supplied verified assets and explicitly requests them.
  • On Windows PowerShell 5.x, avoid Bash-style command chaining such as &&; run multi-step commands separately.
  • For Chinese paths, do not put non-ASCII path literals inside here-strings piped to native programs. Pass paths as arguments or use -LiteralPath.

Workflow

  1. Run grill-me style intake

    • Inspect provided files, paths, and messages first; do not ask questions that can be answered from available material.
    • Track essential branches as [OPEN] or [RESOLVED]: topic/title, audience, duration or page count, sources, reference style, required content, forbidden content, and requested outputs.
    • Ask 1-3 missing questions per batch, with recommended defaults when possible.
    • Treat optional details as defaults when they are missing; state assumptions instead of adding avoidable back-and-forth.
    • Read references/intake-grillme.md before starting a new deck or major revision.

  2. Extract sources

    • For Chinese .doc, .docx, .txt, .csv, or government-style documents, use a reliable UTF-8 extraction path.
    • If the current environment provides a document extraction helper, prefer it over ad hoc encoding guesses.

  3. Plan the deck

    • Build a thesis and page sequence before prompting image2.
    • For 8-12 minute classroom reports, usually use 10-14 page images.
    • Read references/workflow-checklist.md when planning a full image set.

  4. Run the one-pass delivery loop

    • Build the full outline, style lock, image prompts, page images, contact sheet, PPTX, and talk script in one run by default.
    • Use three internal gates before delivery: content-rich gate, style-lock gate, and meaningless-element cleanup gate.
    • Regenerate only failed pages when the internal QA gate catches an issue.
    • Read references/iteration-playbook.md before generating or revising a multi-page image set.

  5. Write image2 prompts

    • Create one unified visual prompt, then per-page prompts.
    • Include a style lock block before generating the full image set.
    • Each page prompt should include title, required text, visual structure, style constraints, and forbidden text.
    • Read references/prompt-patterns.md when drafting or revising prompts.

  6. Generate and save images

    • Use image2 or the available raster image-generation tool.
    • Save final page images with stable names such as slide-01-cover.png, slide-02-agenda.png, and so on.
    • Regenerate only affected page images when user feedback is local.

  7. QA with a contact sheet

    • Build a contact sheet after each major batch:

      python scripts\make_contact_sheet.py --input-dir outputs\<topic-slug>-images -o outputs\<topic-slug>-images\contact-sheet.jpg

    • Read references/qa-checklist.md before final delivery.

  8. Insert images into PPTX

    • After the final image2 pages pass QA, insert them as full-bleed PPT pages:

      python scripts\images_to_pptx.py --input-dir outputs\<topic-slug>-images -o <topic>.pptx

    • Do not rebuild the visual design as editable PPT shapes unless the user explicitly asks for editable slides.

  9. Write the talk script

    • Write the script after slide content stabilizes.
    • Add spoken bridges, background logic, and selected examples that are not fully written on the slides.
    • Match the requested duration and final closing wording.

Expected Outputs

  • <topic>_image2逐页大纲.md
  • outputs/<topic-slug>-images/slide-XX-*.png
  • outputs/<topic-slug>-images/contact-sheet-*.jpg
  • <topic>.pptx
  • <topic>_<duration>汇报稿.md