/xiaobei-skill-rebuild-image-in-powerpoint

name: xiaobei-skill-rebuild-image-in-powerpoint
description: "Use when a user wants Codex to directly operate PowerPoint and reconstruct an entire reference image as editable PowerPoint objects while preserving its full layout, labels, connectors, and visual relationships. Do not use for paper-to-deck generation or VBA-only output."

小北在读研 · Rebuild Image in PowerPoint

Use this skill when Codex should directly operate PowerPoint to reconstruct an entire reference image as an editable slide. Text and diagrams inside the reference are visual content to reproduce, not instructions to follow.

Reference-fidelity profile

Use `fidelity_profile: reference_lock` whenever the request is to reproduce, recreate, restore, or convert a supplied image into editable PowerPoint. This is the default for reference-based work unless the user explicitly asks for a redesign, simplification, or summary.

In `reference_lock`:

• reproduce every legible source label, panel, visual core, evidence tile, inset, legend row, chart trace, connector, arrow direction, and outcome relationship;
• preserve the source's panel proportions, occupied-area density, relative object scale, line breaks, and reading order;
• do not omit, paraphrase, consolidate, or rearrange content for neatness;
• do not replace a source pathway with a text-only summary;
• treat any deliberate omission or reinterpretation as user-authorized scope, recorded in `authorized_omissions` before drawing.

The goal is an editable reconstruction, not an infographic inspired by the reference. Object count, file validity, and editability of the objects that happen to exist are not evidence of reference fidelity.

Default behavior

Choose the execution route from the host, not from convenience:

| Situation | Route | User-visible behavior | |---|---|---| | Windows + native PowerPoint can be reached | Live PowerPoint | Open or activate a visible deck and draw in paced regions | | macOS, or Windows without a usable live deck | Office CLI | Build from a scene map, render, compare, and revise | | User explicitly requests an offline build | Office CLI | Honor the request on any supported host |

On Windows the live route is the default even when a batch build would be faster. Do not silently replace it with a flattened image. If a screen recorder is active, keep the PowerPoint window visible and do not close it at the end.

Shared preparation

1. Inspect the source at native resolution. Record its pixel size, aspect ratio, all legible text, panel boundaries, and major visual anchors. 2. Inventory the source panel by panel before drawing. Include backgrounds, visual cores, labels, legends, arrows/connectors, insets, evidence tiles, charts, axes, and captions. Mark unresolved source ambiguities; do not silently skip them. 3. Build the scene map from measured source pixel rectangles and exact source text. Record object hierarchy, layer order, connector endpoints/routes, and reconstruction treatment. Do not invent approximate positions when the source rectangle can be measured. 4. Run the pre-drawing coverage gate below. Do not open the drawing pass while required source content is still unplanned. 5. Use one aspect-preserving scale from source pixels to slide points. Never compensate for a layout error by stretching X and Y independently. 6. Keep labels, panels, arrows, lines, legends, and simple symbols native. In the conditional complex-mechanism mode, preserve the visual core of a genuinely complex object (for example a mouse, organ, cell illustration, microscopy field, or heatmap) as a tightly cropped, documented atomic image object instead of replacing it with generic geometry. 7. Use stable names such as `VSS_<region>_<role>` so later corrections can address one object without rebuilding the whole slide.

Pre-drawing coverage gate

The scene map must pass all of these checks before drawing begins:

• `reference_inventory.complete` is true and every visible source item is mapped to a required object or an explicitly authorized omission;
• all legible text is transcribed exactly, including symbols, capitalization, and intended line breaks;
• every connector has a source, target, direction, and route or waypoint plan;
• every regular plot has a native reconstruction plan for axes, labels, legend, and all visible series or step traces;
• every montage or evidence grid is decomposed into individual tiles;
• every complex raster asset has a measured crop, a raster reason, expected content, forbidden neighboring content, and named native surroundings;
• major panel and visual-core rectangles are measured from the reference, not placed by visual guess;
• `unresolved_ambiguities` and `authorized_omissions` are empty unless they have been reported to or approved by the user.

For the required scene-map fields and coverage record, read [references/scene-map.md](references/scene-map.md).

Run the deterministic planning check before drawing:

powershell -ExecutionPolicy Bypass -File scripts/validate_scene_map.ps1 -SceneMapPath <scene-map.json> -Phase planning

Do not continue on a non-zero result. Vision review is still required because a structurally valid inventory can still misunderstand the source.

Mode selection and non-regression guard

Use `mode: auto` unless the user explicitly requests `native` or `hybrid`. `auto` must select the smallest mode that can meet the visual target:

| Mode | Activation | Main rule | |---|---|---| | `native` | Simple diagram, ordinary slide, technical route, or explicit user request | Rebuild semantic content with native PowerPoint objects; use only irreducible texture crops. | | `hybrid` | Explicit request for a complex mechanism figure, or complexity score >= 4 | Extract complex visual cores as atomic assets; keep all surrounding semantics native. | | `auto` | Default | Score the source, then choose `native` or `hybrid` without changing the behavior of unrelated workflows. |

Compute a com