chart-data-extractor

Chart Data Extractor Skill

Extracts data from images of charts and graphs — bar charts, line charts, pie charts, scatter plots, and tables in images — producing a structured data table that can be used in spreadsheets or rebuilt in any charting tool. Built to leverage Opus 4.7 pixel-level image analysis capabilities.

Required Inputs

Ask the user for these if not provided:

• The chart image (upload a screenshot or image file)
• Chart type (if ambiguous — bar / line / pie / scatter / other)
• What matters most (approximate trends / precise values / specific data points / categorisation)
• Known axis values (optional — if the user knows the max/min values to anchor the extraction)

Output Structure

1. Chart Identification

| Attribute | Value |

|---|---|

| Chart type | [Bar / Line / Pie / Scatter / Area / Other] |

| Chart title (if visible) | [Title text] |

| X-axis label | [Label + unit] |

| Y-axis label | [Label + unit] |

| Number of series | N |

| Legend categories | [List] |

| Data period (if time-based) | [Start — End] |

2. Extracted Data Table

| [X axis] | [Series 1] | [Series 2] | ... |

|---|---|---|---|

| [Value] | [Value] | [Value] | |

3. Confidence Levels

For each data point or series, flag confidence:

• High confidence: data points where the value is clearly readable against gridlines or labels
• Medium confidence: data points where the value is interpolated between gridlines
• Low confidence: data points where the value is ambiguous or overlaps with other elements

Low-confidence points should be explicitly listed — not silently included in the main table.

4. Notable Observations

Observations that the data itself reveals:

• Peak value: [Value, when, in which series]
• Lowest value: [Value, when, in which series]
• Largest delta between series: [Details]
• Any anomalies or outliers visible in the chart

5. Reconstructed Source

CSV format for direct use:

[x_axis],[series_1],[series_2]
[value],[value],[value]

6. Assumptions and Caveats

• Grid resolution: [How precisely values could be read — e.g. "Y-axis has major gridlines every 10 units, minor every 2"]
• Interpolation used: [Any values that required estimating between gridlines]
• Unclear data: [Anything in the chart that could not be read reliably]
• Axis scale: [Linear/logarithmic/etc — note if not obvious]

7. Follow-up Options

Ask the user which of these they want:

• Rebuild the chart in a specified format (Excel formula, Python matplotlib, D3, etc.)
• Produce a narrative description of what the chart shows
• Compare this data against another chart or source
• Flag potentially misleading visual choices in the original (truncated axes, misleading scales, etc.)

Quality Checks

• [ ] Every extracted number specifies which series it belongs to
• [ ] Confidence levels are explicit for ambiguous points
• [ ] Low-confidence values are flagged separately, not silently included
• [ ] Assumptions about axis scale and interpolation are stated
• [ ] CSV output is clean and directly usable

Anti-Patterns

• [ ] Do not silently include low-confidence data points in the main table — flag them separately so the user knows which values to verify
• [ ] Do not assume a linear scale without confirming it — logarithmic axes make extracted values incorrect by orders of magnitude if misread
• [ ] Do not report extracted values with false precision — if the chart's Y-axis only shows gridlines every 10 units, a reported value of 37 is invented, not extracted
• [ ] Do not omit the assumptions and caveats section — partial image quality, overlapping bars, or unlabelled axes must be disclosed

Example Trigger Phrases

• "Extract the data from this chart"
• "Transcribe the numbers in this graph"
• "Turn this chart image into a spreadsheet"
• "Digitise this chart so I can rebuild it"
• "What are the exact values in this bar chart?"

Why This Works Better on Opus 4.7

Earlier models struggled with pixel-level data transcription from charts, often hallucinating values or misreading gridline positions. Opus 4.7 uses a higher image resolution (2576px vs 1568px) with coordinates mapping 1:1 to pixels, making chart data extraction reliable for practical use.