Happy Figure Edit

Happy Figure Edit

Convert one input image into a faithful, editable output.svg, then export output.pptx only after the SVG passes review.

You are the reconstruction expert. The runner prepares evidence, validates your JSON, materializes cropped assets, renders diffs, and converts SVG to PPTX. It does not replace your visual judgment.

Setup Check

Run commands from the directory containing this SKILL.md. If .venv/bin/python is missing, install once:

python3 -m venv .venv
.venv/bin/python -m pip install -r requirements.txt
brew install resvg

Use .venv/bin/python for all scripts. resvg is used only for rendered.png, diff.png, and the pixel-diff section of quality_report.json; SVG/PPTX generation still works if it is missing.

Main Workflow

All normal outputs must stay under the skill directory:

• Work directory: outputs/<image-stem>_work
• Delivery directory: outputs/<image-stem>

Do not write reconstruction outputs to /tmp, ~/Documents, or the caller's current working directory unless the user explicitly asks for that path.

Execution discipline:

• Keep the shell working directory at the directory containing this SKILL.md for every runner command.
• Do not cd into outputs/<image-stem>_work and then repair relative paths by trial and error.
• Use only .venv/bin/python scripts/run_expert_mvp.py ... from the skill root for normal commands.
• Do not inspect scripts/, _vendor/, or converter source during a normal reconstruction. Read source code only when a runner command fails or the user asks to debug the skill itself.
• If a temporary helper script is useful for writing expert_response.json, put it under outputs/<image-stem>_work/, but execute it from the skill root, for example .venv/bin/python outputs/<image-stem>_work/build_expert_response.py. Do not package helper scripts as deliverables.

1. Prepare a run directory:

   .venv/bin/python scripts/run_expert_mvp.py \
     --image <absolute-image-path> \
     --out-dir outputs/<image-stem>_work
   

   This writes evidence.json, a conservative baseline element_analysis.json, fallback output.svg, report.html, quality_report.json, and quality_summary.txt. It also writes a baseline element_overlay.png, but that overlay usually contains only the full-canvas fallback box and is not useful for visual decomposition. It does not generate PPTX during preparation.

2. Read the required inputs before writing anything:

   • Open the original image with view_image at high detail.
   • Read outputs/<image-stem>_work/evidence.json for exact canvas width and height.
   • Read prompts/expert_structure.md for element decomposition and asset strategy rules.
   • Read prompts/expert_svg.md for SVG constraints.
   • Do not open the initial baseline element_overlay.png when it only shows the conservative full-canvas fallback. Use the original image for decomposition.
   • Do not read runner or converter implementation files unless the normal workflow errors.

3. Decompose the figure by eye:

   • Identify titles, labels, section containers, panels, tables, axes, nodes, arrows, legends, callouts, icons, logos, screenshots, photos, heatmaps, and dense plots.
   • Split every icon, logo, badge, button, avatar, app/document glyph, and complex small symbol into its own tight crop_nobg element unless it is only a primitive mark such as a plain circle or line.
   • Estimate each element bbox as integer pixels [x, y, w, h] in the original canvas coordinate system.
   • Keep text, simple shapes, arrows, tables, axes, formulas, and panels editable whenever they can be faithfully redrawn.

4. Write outputs/<image-stem>_work/expert_response.json manually using the schema below. The SVG must reference cropped rasters as assets/<box_id>.png; apply-response creates those files from the element bbox.

5. Apply and review SVG:

   .venv/bin/python scripts/run_expert_mvp.py apply-response \
     --run-dir outputs/<image-stem>_work \
     --response outputs/<image-stem>_work/expert_response.json
   

   During apply-response, the runner materializes assets, validates output.svg, renders it to rendered.png, computes diff.png, writes review tiles, and performs one deterministic crop_nobg boundary repair pass. It does not generate PPTX. Iterate on SVG quality first; PPTX export happens only in package-run after the final SVG review is acceptable.

6. Verify visually:

   • Open the original image, the regenerated outputs/<image-stem>_work/element_overlay.png, outputs/<image-stem>_work/rendered.png, and outputs/<image-stem>_work/diff.png.
   • Read outputs/<image-stem>_work/quality_summary.txt first. Use quality_report.json only when the summary is not enough.
   • Inspect asset_integrity_review in quality_summary.txt first. Open the listed asset_review_tiles/ files before generic review_tiles/; these are dedicated checks for clipped icons, missing card backgrounds, and bad crop_nobg bboxes.
   • Inspect at most the top five files listed under outputs/<image-stem>_work/review_tiles/. These local original/rendered/diff tiles are for high-value checks, not for exhaustively reviewing every diff.
   • Prioritize asset_integrity_review, then granularity_warnings, then review tiles before global pixel diff scores. Large text/group/panel bboxes can hide missing icons, missing card backgrounds, and wrong connectors.
   • Review in this order: incomplete crop/crop_nobg assets; missing icon/card/container backgrounds; graph node shape and edge topology mistakes; missing characters/content; small text/icon offsets.
   • Prioritize declared_region_review for high-diff content hidden inside broad zones/panels, unassigned_diff_regions for missing glyphs/unmodeled elements, and diff_regions[].near_crop_edges for incomplete crop or crop_nobg bboxes.
   • Read outputs/<image-stem>_work/crop_repair_report.json; accept applied repairs unless visual review shows overreach, and inspect review candidates before manually changing bbox or strategy.
   • Iterate first on incomplete crops, missing card backgrounds, wrong graph shapes/links, missing characters, and unmodeled elements. Treat tiny text/icon offsets as acceptable unless they materially change readability or meaning.
   • Re-run apply-response from the skill root after each expert_response.json update. Do not run commands from inside the work directory.

7. Package the run for delivery instead of manually copying selected files:

   .venv/bin/python scripts/run_expert_mvp.py package-run \
     --run-dir outputs/<image-stem>_work \
     --out-dir outputs/<image-stem>
   

   This generates the final output.pptx from the reviewed output.svg, then preserves the canonical filenames used by report.html: output.svg, output.pptx, element_overlay.png, rendered.png, diff.png, quality_summary.txt, quality_report.json, crop_repair_report.json, run_status.json, review_tiles/, and assets/. It also derives a Figma delivery under figma/: figma/output.figma.svg embeds assets/* as base64 data URIs so it can be dragged into Figma, and figma/figma_payload.json bundles the canonical SVG, element analysis, and base64 assets for a future importer plugin. The canonical output.svg keeps its relative assets/<box_id>.png hrefs and is not modified. Do not create renamed copies of output.svg or output.pptx unless the user explicitly asks for duplicate filenames.

Asset Strategy Rules

Choose exactly one strategy per element:

Decision order:

1. If SVG primitives can redraw it faithfully, use svg_self_draw.
2. If redraw would distort it but the foreground can be separated, use crop_nobg with a tight bbox around the object only.
3. If fidelity matters and the background must stay, use crop.
4. Do not simplify these foreground objects into editable line art and do not hide them inside a larger svg_self_draw panel.
5. Do not make one crop larger than 75% of the canvas; keep decomposing the figure.
6. Do not use one giant text, group, connector, or panel bbox to cover many internal labels/icons; split by semantic element so review artifacts can expose missing pieces.
7. For icon cards, separate the editable SVG card/container from the crop_nobg foreground icon, or use crop when the card background must remain raster-faithful.
8. For network/KG regions, preserve node shape categories and edge topology explicitly; square nodes must not become circles unless the source uses circles.

expert_response.json Contract

The response must validate against happyfigure.edit.expert_response.v1:

{
  "schema": "happyfigure.edit.expert_response.v1",
  "element_analysis": {
    "schema": "happyfigure.edit.element_analysis.v1",
    "source": "skill_expert_agent",
    "canvas": {"width": 1376, "height": 768},
    "strategy_summary": "Short decomposition summary.",
    "elements": [
      {
        "box_id": "T001",
        "source_candidate_ids": ["T001"],
        "bbox": [10, 20, 120, 32],
        "kind": "text",
        "asset_strategy": "svg_self_draw",
        "confidence": "high",
        "reason": "Readable title redrawn as editable SVG text.",
        "evidence": ["Visible title at top left."]
      }
    ],
    "review": {"status": "none", "notable_adjustments": []}
  },
  "svg": "<svg xmlns=\"http://www.w3.org/2000/svg\" width=\"1376\" height=\"768\" viewBox=\"0 0 1376 768\">...</svg>",
  "notes": ["Short implementation notes."]
}

Validation constraints:

• canvas.width and canvas.height must match evidence.json.
• Every bbox must be four integers inside the canvas.
• Every element must use one of svg_self_draw, crop, or crop_nobg.
• SVG root width, height, and viewBox must match the canvas.
• SVG must not contain file://, data:, base64, external URLs, <style>, <filter>, <mask>, <clipPath>, <foreignObject>, <textPath>, <symbol>, or <use>.
• Raster <image> href values must be relative assets/<box_id>.png paths.

Mainline Boundaries

Use only scripts/run_expert_mvp.py, scripts/run_expert_mvp.py apply-response, and scripts/run_expert_mvp.py package-run for normal reconstruction and delivery.

Do not describe fallback outputs as final high-quality editable reconstructions.