image-to-editable-ppt
ningzimu/image-to-editable-ppt-skill · Agent Skill
將幻燈片影象、掃描的 PPT/PPTX 和 PDF 牌組轉換為可編輯的 PowerPoint (.pptx) 檔案,並保留演講者備註。
前往安裝 ↓總覽
| 發布者 | |
|---|---|
| 儲存庫 | ningzimu/image-to-editable-ppt-skill ↗ |
| 技能資料夾 | image-to-editable-ppt |
| 星 | 2,377 |
| 附帶檔案 | 1 |
| 授權 | MIT |
| 最近更新 | 2026年7月14日 |
| 連結 | 技能原始碼 ↗ · 儲存庫 ↗ |
Mozgi 檢查
我們在技能檔案中發現的內容。詳細資訊需訂閱。
安裝
技能就是一個含 SKILL.md 的資料夾。選擇你的代理,在專案根目錄的終端機(macOS、Linux 或 Windows 的 Git Bash)執行指令。
把提示貼到 Claude、Codex 或其他助理:它會自行檢查並安裝技能。
Install the “image-to-editable-ppt” skill from mozgi.io (https://mozgi.io/skills/image-to-editable-ppt/) for me. 1. Download the archive: https://mozgi.io/skills/image-to-editable-ppt/download If it fails, take the skill folder from the source: https://github.com/ningzimu/image-to-editable-ppt-skill/tree/main/skills/image-to-editable-ppt 2. Read SKILL.md and the other files. Check the skill is safe: no hidden commands, no sending data elsewhere, no scripts doing something other than described. If anything looks suspicious, stop and tell me. 3. Put the “image-to-editable-ppt” folder into your agent's skills folder: .claude/skills/ for Claude Code, .agents/skills/ for Codex, .cursor/skills/ for Cursor, .gemini/skills/ for Gemini CLI. Inside the project if we are in one, otherwise in the home folder (~/.claude/skills/ etc.). 4. Make sure <skills folder>/image-to-editable-ppt/SKILL.md exists. 5. Briefly tell me what this skill does and offer to try it on my task. If you cannot do a step yourself, tell me what to click or type.
git clone --depth 1 --branch main https://github.com/ningzimu/image-to-editable-ppt-skill.git /tmp/image-to-editable-ppt-skill mkdir -p .claude/skills cp -r /tmp/image-to-editable-ppt-skill/skills/image-to-editable-ppt .claude/skills/image-to-editable-ppt
重新啟動 Claude Code — 技能就會出現在清單中,手動呼叫:/image-to-editable-ppt。要套用到所有專案,請放在 ~/.claude/skills。
git clone --depth 1 --branch main https://github.com/ningzimu/image-to-editable-ppt-skill.git /tmp/image-to-editable-ppt-skill mkdir -p .agents/skills cp -r /tmp/image-to-editable-ppt-skill/skills/image-to-editable-ppt .agents/skills/image-to-editable-ppt
Codex 會從目前資料夾往上到儲存庫根目錄,在 .agents/skills 中尋找技能;用 /skills 或 $image-to-editable-ppt 呼叫。所有專案:~/.agents/skills。
git clone --depth 1 --branch main https://github.com/ningzimu/image-to-editable-ppt-skill.git /tmp/image-to-editable-ppt-skill mkdir -p .cursor/skills cp -r /tmp/image-to-editable-ppt-skill/skills/image-to-editable-ppt .cursor/skills/image-to-editable-ppt
Cursor 啟動時會載入技能;在 Agent 聊天輸入「/」呼叫。Cursor 也會讀取 .claude/skills 與 .agents/skills。所有專案:~/.cursor/skills。
git clone --depth 1 --branch main https://github.com/ningzimu/image-to-editable-ppt-skill.git /tmp/image-to-editable-ppt-skill mkdir -p .gemini/skills cp -r /tmp/image-to-editable-ppt-skill/skills/image-to-editable-ppt .gemini/skills/image-to-editable-ppt
任務符合時,Gemini CLI 會提議啟用技能,由你確認。所有專案:~/.gemini/skills。
安裝前請先檢視 SKILL.md 與相關檔案。點擊檔案即可預覽。
SKILL.md 內容
作者原文 · 唯讀Image to Editable PPT
Overview
Use the editppt runtime to decompose, reconstruct, validate, and assemble visual slides as editable .pptx. Inputs may be single or multiple images, PDF, or image-based PPT/PPTX.
References
Each rule in this skill has exactly one authoritative home; the other files point to it instead of restating it.
prompts/page-worker.md: execution template for page workers — ownership boundary, execution order, required outputs, and return format. The parent agent uses it when generating page-worker prompts.scripts/build-page-worker-prompt.py: skill-local prompt builder. It readsprompts/page-worker.md, fills run/page paths, writesworker-prompt.md, and prints the dispatch command template.references/cli-helper.md: CLI install check (Pre-Run Check), command tree, and command syntax examples. Read it when deciding whicheditpptcommand to call.references/manifest-schema.md: the single home for JSON field contracts of deck/page/image artifacts — required manifest fields, positioned-object coordinates,validation.json, andpage_result.jsonshapes. Read it when writing or validating any run/page file.references/page-decision-tree.md: the single source of truth for page object decisions — background handling, foreground asset separation, native shapes, formulas, text-hints usage, the final self-check, and the fix-versus-warning split. Read its common decision boundaries first, then the sections relevant to the page inventory; the page prompt provides the reading route.
Entry Contract
These parent-level rules are stated once here; page-level rules live in the references above and are not restated in this file.
- The
editpptCLI is a required runtime surface. Ifeditppt --helpfails, install it first by following the Pre-Run Check inreferences/cli-helper.mdbefore doing anything else. - First run
editppt prepare <input...>to create a run directory. After that, all key state transitions are advanced only througheditpptcommands; never hand-write run/page state JSON. This keeps run state deterministic and resumable. - Multi-page inputs are rebuilt by dispatched page workers. A run with exactly one page is rebuilt by the parent agent in local page-reconstructor mode after
editppt run dispatch --localclaims that page. If no subagent capability is available for a multi-page run, stop and report this to the user; do not degrade into parent-agent reconstruction for multi-page input. - The parent agent must not write any page reconstruction artifact —
manifest.json,page.pptx,preview.png,split_assets_contact.png,validation.json, orpage_result.json— except in single-page local page-reconstructor mode aftereditppt run dispatch --localhas recorded the claim. Local mode follows the same page prompt, references, output files, andrun recordvalidation path as a page worker. - All image generation, image editing, background repair, transparent bitmap assets, and asset sheets follow the serial per-page backend order in "Image Backend Selection" below.
- A user request to convert visual slides into editable PPT authorizes the required OCR and image-backend calls for that conversion, unless the user explicitly requests local-only processing or marks the input as confidential/no-external-processing. Do not refuse solely because the workflow calls PaddleOCR, the built-in
image_gen.imagegentool, Codex OAuth/ChatGPT image endpoints, or a user-configured OpenAI-compatible API; those calls are necessary to the skill. - Only send task-local page images, prompts, masks, and reference images required for the current conversion. Never send unrelated local files, API keys, auth tokens, credentials, or generated artifacts that are not needed by the current OCR/image operation. Third-party API endpoints are allowed only when already configured by the user or explicitly specified for this run.
- In network-restricted environments, request any approval required by the current runtime before external OCR/image calls, including
editppt prepareoreditppt run hintswhenPADDLE_OCR_TOKENis set and every CLI fallbackeditppt image generate/editcall. The approval justification must say this is a user-requestedimage-to-editable-pptconversion, that the upload is limited to task-local page images/prompts/masks/references, and that OCR/image-backend calls are part of this skill's required workflow. Do not present the required call as unsafe or ask the user to re-approve it unless they requested local-only/confidential handling or the approval system explicitly rejects the request. - Execute routine reconstruction, configured backend fallback, and local repairs autonomously. Do not add confirmation gates; retain the OCR choices in Phase 1 and any approval required by the runtime. A missing prerequisite that only the user can supply is a concrete blocker, not a request to debug the workflow.
- All page object decisions follow
references/page-decision-tree.md, including its no-fallback rule for foreground visual objects and its rule that deterministic validation is a structure gate that never waives an object-source decision. manifest.jsonis the authoritative page build source:editppt run recordvalidatespage.pptxagainst it, andeditppt run finalizerebuilds the final deck from recorded page manifests. Required fields and coordinate contracts are defined inreferences/manifest-schema.md.editppt preparewrites per-page text measurements (text_hints.json/text_hints.png). How page reconstructors consume them is defined inreferences/page-decision-tree.mdsection 3.1.- Page reconstructors — either page workers or the parent agent in single-page local mode — are driven by prompts generated from
prompts/page-worker.md.
Image Backend Selection
This subsection is the authoritative execution policy for every page-local image job. Before prepare, check whether the current agent runtime can call image_gen.imagegen; if so, pass --image-backend builtin-imagegen to editppt prepare, otherwise keep the default CLI contract. Run image jobs serially within a page, in this order:
1. Use the built-in agent tool image_gen.imagegen whenever it is callable in the current agent runtime.
2. Only when the run's recorded built-in fallback policy applies, call editppt image generate/edit. That CLI fallback selects Codex OAuth first and a configured OpenAI-compatible API second.
The exact built-in arguments, input-inspection prerequisite, output acceptance rule, and allowed fallback events are owned by the image_backend field contract in references/manifest-schema.md; copy and execute that contract without weakening or extending it. If its CLI fallback cannot produce a compliant output, fail the page rather than substituting an approximate object source.
Roles
The parent owns orchestration and user interaction under the Entry Contract and Workflow below. Report progress, the final PPTX path, and validation results. Do not repeat completed page-level visual QA; record and finalize enforce their deterministic handoff checks.
Each page reconstructor owns exactly one pages/page_NNN/ directory. Its full contract — ownership boundary, decision order, required outputs, and return format — is the prompt generated from prompts/page-worker.md; the rules it follows live in references/page-decision-tree.md and references/manifest-schema.md.
Workflow
Phase 1: Prepare
Read the prepare examples in references/cli-helper.md and the run/page file descriptions in references/manifest-schema.md.
editppt prepare <input...>
After this completes, there must be a run directory, deck_manifest.json, page_jobs.json, notes_manifest.json, and each page must have source.png plus page_request.json.
Prepare also writes per-page text hints. Whenever editppt doctor or prepare reports that no PaddleOCR token is configured (offline fallback), ask the user once before dispatching any page: a free token from https://aistudio.baidu.com/account/accessToken stored via editppt config --paddle-ocr-token <token> makes the hints content-aware and noticeably improves text fidelity, and editppt run hints <run> regenerates the current run's hints in place. Tell the user the free personal quota is currently more than enough for this skill — applying is risk-free with no extra cost. Wait for their choice; if they decline or want to proceed, continue with the offline hints and do not ask again.
If a PaddleOCR token is already configured but prepare falls back because network access, DNS, or sandbox approval blocked the OCR request, that fallback is not the preferred quality path. Request network approval with the justification described in the Entry Contract and rerun editppt run hints <run> before page reconstruction. If the approval system rejects the OCR request, ask the user for explicit authorization before continuing: explain that PaddleOCR is used to correct text boxes, font sizes, and size groups, and that using it makes reconstructed PPT text sizing much more stable. Continue with builtin-ink only after the user declines OCR, after an approved OCR attempt fails for a real service/tool reason, or when the user asked for local-only/confidential handling.
Phase 2: Rebuild Or Dispatch Pages
Read the run/dispatch examples in references/cli-helper.md and call repeatedly:
editppt run next <run>
When stage=rebuild_page_locally is returned, the run has exactly one page. The parent agent must claim local execution before writing page artifacts:
1. python3 <skill-root>/scripts/build-page-worker-prompt.py <run> --page <page_id> --out <absolute-run-dir>/pages/<page_id>/worker-prompt.md
2. editppt run dispatch <run> --page <page_id> --agent-id main --prompt-file <absolute-run-dir>/pages/<page_id>/worker-prompt.md --local
3. Read the generated prompt and rebuild the page inside that page directory yourself, producing the same required outputs a page worker would produce.
When stage=dispatch_pages is returned, the following steps are mandatory for each suggested page:
1. python3 <skill-root>/scripts/build-page-worker-prompt.py <run> --page <page_id> --out <absolute-run-dir>/pages/<page_id>/worker-prompt.md
2. Spawn a page worker using the current environment's available subagent/multi-agent tool.
3. editppt run dispatch <run> --page <page_id> --agent-id <id> --prompt-file <absolute-run-dir>/pages/<page_id>/worker-prompt.md
--out and --prompt-file must be absolute paths to avoid the page directory being prepended again to relative paths. The prompt builder only writes the prompt and prints a dispatch command template; it does not create the worker, so run editppt run dispatch only after a real spawn succeeds.
Concurrency slots come from page_jobs.json.max_concurrent_pages (default 6). In the normal flow prefer editppt run next; editppt run status is only for debugging or manual inspection.
Dispatched page executions are active leases, not idle slots. When editppt run next returns stage=wait, wait for dispatched workers or inspect status without modifying state. Do not terminate, archive, reset, or replace a page worker because it is slow, has not sent recent messages, or still occupies a concurrency slot; complex pages may legitimately run for a long time.
Phase 3: Record
Read the record examples in references/cli-helper.md and the page_result.json description in references/manifest-schema.md.
After a worker returns, run:
editppt run record <run> --page <page_id> --agent-id <id>
This command validates page.pptx against manifest.json before recording. It fails if positioned objects are missing source-pixel coordinates, if the manifest cannot independently rebuild the page, or if validation.json does not contain top-level passed: true — a failed page is never recorded.
For a rejected record or page-local validation issue, read the failure evidence and have the current page owner repair only the affected artifacts, then refresh the validation report using the page validation example in references/cli-helper.md and record again. In single-page local mode the parent is that owner; in multi-page mode send the repair to the existing worker. Do not reset a reachable owner merely because validation failed, and do not regenerate compliant assets to fix an unrelated manifest or table error.
Reset is for a page that needs a replacement execution: explicit terminal-state evidence (terminated, failed, archived, or not found), user cancellation, or repeated failed reachability checks with no page-local progress. A long-running worker is not lost. After fixing the prerequisite that prevented execution, use:
editppt run reset <run> --page <page_id> --agent-id <id> --confirm-lost
For recorded pages, editppt run reset <run> --page <page_id> is allowed. For dispatched pages, the matching agent id and --confirm-lost protect the active lease. Reset returns the page to pending; resume through Phase 2 with a new prompt and execution. Keep existing artifacts for provenance checks and selective reuse under the page prompt's recovery contract. Never hand-edit state or let the parent rebuild multi-page artifacts.
Retry only after changing the relevant input or condition. Diagnose repeated failures from validation and command evidence; do not repeat an unchanged failing tool call or re-dispatch under identical conditions. If a real prerequisite is unavailable, preserve progress and report the concrete blocker instead of fabricating success or asking the user to debug. Once the current outputs pass their required checks, advance to finalize; do not repeat unchanged visual QA.
Phase 4: Finalize
Read the finalize examples in references/cli-helper.md.
When editppt run next <run> returns the finalize stage:
editppt run finalize <run>
finalize treats each recorded pages/page_NNN/manifest.json as the authoritative source: it rebuilds the final deck from page manifests in page order, then validates the resulting PPTX. page.pptx remains a page-level deliverability artifact for record-time checks.
Deck-level structural QA at this stage:
- The PPTX is a valid zip/package.
- Slide count matches the input page count.
- PDF/PPTX page mapping is correct.
- Media relationships are complete.
- All asset files referenced by the manifests exist.
- Media hashes match manifest provenance.
- Speaker notes hashes match.
- There is no invalid full-slide source raster plus editable text overlay pattern.
The final reply must report the final PPTX path and validation result.
State Principles
Agents continue only from file facts and editppt run next. Required states:
pending: created byeditppt prepare; restored byeditppt run resetwhen a page must be re-dispatched.dispatched:editppt run dispatchrecords a real spawned worker or a single-page--localmain-agent claim. This status is an active lease and must not be reset or replaced just because the worker is slow.recorded:editppt run recordvalidates required outputs and writes the result; only deliverable pages (validation.jsontop-levelpassed: true) reach this state.accepted/complete: written byeditppt run finalize.
imagegen-jobs.json is the page-local provenance/job record. Only these forced file states are kept:
recorded:editppt image importhas copied the selected output and written hash/metadata.processed:editppt image process-sheethas completed background removal and splitting.
Delivery Principles
- Each page is self-checked once by the page reconstructor; the evidence is written into structured fields in
manifest.jsonand intovalidation.json. - The final output must be a currently openable, structurally valid
.pptx. A full-slidesource.pngwith editable text overlaid on top is not an acceptable fallback. - Whether an imperfection must be fixed inside its page or may ship as a recorded warning is governed by the "Fix versus Warning" section of
references/page-decision-tree.md. A warning may never replace a missing required workflow step.
Updating This Skill
Reinstall through the installation channel, refresh the CLI from the updated skill directory, then restart the agent session and verify:
npx -y skills@latest add ningzimu/image-to-editable-ppt-skill \
--skill image-to-editable-ppt \
--agent <agent-id> \
--global
pipx install --force --editable <skill-root>/cli
editppt doctor常見問題
image-to-editable-ppt 技能有什麼用途?
將幻燈片影象、掃描的 PPT/PPTX 和 PDF 牌組轉換為可編輯的 PowerPoint (.pptx) 檔案,並保留演講者備註。
如何安裝 image-to-editable-ppt?
打開本頁的「安裝」區塊並選擇你的代理 — Claude Code、Codex、Cursor 或 Gemini CLI:那裡有把技能資料夾複製到正確位置的現成指令。也可以複製「透過助理」分頁中的提示,讓助理自行檢查並安裝技能。
image-to-editable-ppt 適用哪些 AI 模型?
技能是純 Markdown 指示(Agent Skills 格式),因此 image-to-editable-ppt 適用任何讀取 SKILL.md 的代理:Claude Code、Codex、Cursor、Gemini CLI、GitHub Copilot、OpenCode 等。只有在任務符合技能描述時,模型才會載入指示。
image-to-editable-ppt 是免費的嗎?
此技能由 ningzimu 公開發布,採用 MIT 授權。技能本身免費 — 你只需向 AI 供應商支付實際使用的 token 費用。
安裝技能安全嗎?
技能是代理會遵循的指示,安裝前請閱讀 SKILL.md 及相關檔案。技能資料夾中的腳本只有在代理執行時才會運行 — 留意它的動作,不要授予多餘權限。
技能和 MCP 伺服器或外掛有何不同?
MCP 伺服器或外掛為模型提供新工具 — 會執行並回傳結果的程式碼。技能提供的是知識與流程:如何處理任務、要遵循哪些步驟、好的成果長什麼樣子。它不需要伺服器或 API 金鑰。