| name | chart-revive |
| description | Turn chart screenshots, chart images, and chart regions from PDFs or slides into native editable PowerPoint charts with embedded data, CSV export, confidence labels, and an offline audit report. Use when a user wants to recreate, recover, reverse-engineer, restyle, translate, or update a static bar, column, stacked-column, line, pie, or doughnut chart instead of keeping it as a flat image. |
Chart Revive
Rebuild a static chart as a real PowerPoint chart. Keep every recovered value reviewable; never present pixel estimates as source truth.
Workflow
- Confirm the input contains a supported 2D chart: column, stacked column, horizontal bar, line, pie, or doughnut.
- Reject unsupported charts rather than forcing a conversion. This version does not support negative values, combo, dual-axis, scatter, bubble, radar, waterfall, 3D, pictorial, or map charts.
- Create a task-local output directory. Do not modify the source image, PDF, slide deck, or an existing PowerPoint file.
- Read references/extraction-playbook.md before extracting data from pixels.
- Read references/manifest-schema.md, then write
chart-manifest.json beside a PNG, JPEG, or WebP copy of the source chart.
- Validate the manifest before building:
python <skill-dir>/scripts/chart_revive_cli.py validate <workdir>/chart-manifest.json
- Build the editable chart and audit artifacts:
python <skill-dir>/scripts/chart_revive_cli.py build <workdir>/chart-manifest.json --output <workdir>/result --hero
- Read references/qa-checklist.md. Inspect
preview.png, report.html, summary.json, and the native .pptx structure. If a slide renderer is available, render the PowerPoint and compare it with the source.
- Correct the manifest, not the generated files. Rebuild after each correction. Stop after three iterations unless the user explicitly requests more.
- Deliver
revived-chart.pptx, revived-chart.csv, report.html, preview.png, summary.json, and the manifest. State which values still need verification.
Confidence Rules
Assign confidence per data point:
1.00: supplied by the user or printed as an unambiguous data label.0.90-0.99: directly readable from a clear labeled axis or table.0.70-0.89: interpolated between clear ticks.<0.70: rough visual estimate, occluded mark, or ambiguous scale. Flag it for manual verification.
Use source.method honestly:
provided: all values came from user-provided data.axis-read: values were read from labels or calibrated axes.visual-estimate: values were estimated from pixels.mixed: more than one method was used.
Do not raise confidence merely because the reconstructed preview looks similar.
Runtime
The deterministic builder requires Python 3.10+, Pillow, and python-pptx. Check imports before running. If dependencies are missing, explain what is missing and obtain permission before installing from scripts/requirements.txt.
The CLI does not call an OCR service, image API, or external server. Image understanding is performed by the active multimodal agent; the CLI validates the manifest and creates files locally.
Boundaries
- Do not claim automatic or pixel-perfect numeric recovery.
- Do not invent hidden categories, truncated labels, units, baselines, or series.
- Do not use the result as authoritative evidence until low-confidence values are checked.
- Do not reproduce a chart when the user lacks permission to use the source.
- Do not add macros, external workbook links, remote fonts, CDN assets, or network calls.
- Prefer an honest partial result over a confident fabrication.
