TL;DR
A good chart caption names the pattern and the implication, not the bars. The fast way to get one: feed a vision model (ChatGPT GPT-5.5, Claude Opus 4.7, or Gemini 3.1 Pro, all available as of June 2026) the chart image plus the key numbers typed out, then ask for exactly two sentences. The image gives the model the shape; the typed numbers stop it from inventing figures. Budget 30 seconds per chart once your prompt is dialed in.
The task
You are building a deck or a dashboard and every chart needs a caption. The weak version describes what the chart shows: “revenue grew 12% Q2 over Q1.” The strong version tells the audience why it matters: “revenue is growing but margin is flat, so pricing is the next conversation.” The European Union’s data-visualisation guidance makes the same point: a title should state the main message in ten words or less, not restate the axis labels (data.europa.eu).
You want the strong version, consistently, at 30 seconds per chart.
When AI is the right tool
- You have 5 to 30 charts to caption in a single deck and want one consistent voice.
- The chart is legible: a trend line, bar chart, stacked bar, or simple two-series comparison.
- You can state the key numbers (start, end, peak, dip) yourself.
When not to lean on AI alone: dense multi-axis or small-multiple charts where the model cannot tell which series maps to which axis; board-level decks where one wrong takeaway burns credibility; and regulated reporting (financial filings, clinical results) where caption wording carries compliance weight. In those cases, draft with AI and verify every number by hand.
Feed it the image AND the numbers
Modern vision models read charts far better than they did a year ago, but they still misread axes and occasionally fabricate a figure. The fix is cheap: paste the screenshot and type the handful of numbers that matter. The image anchors the shape; your typed numbers anchor the math.
What goes in the prompt:
- The chart image (screenshot or export).
- Chart type, x-axis, y-axis with units, and the series names.
- The numbers you care about: first value, last value, any peak or dip, and the % change you want quoted.
- Audience: who reads the deck and what decision they own.
- Your one-line hypothesis, if you have one.
Which model to use (June 2026)
All three flagships accept image uploads and can caption a chart. Differences that matter for this task:
| Model | Image input | Notable for charts | In-app context |
|---|---|---|---|
| ChatGPT GPT-5.5 | Yes (high-detail patches) | Strong on dense, high-resolution dashboards and small labels | ~320 pages on Plus; full 1M on $200 Pro |
| Claude Opus 4.7 | Yes (up to ~2,576 px long edge) | Best at not fabricating; reports missing data instead of guessing; large CharXiv figure-reading gain over 4.6 | 1M tokens |
| Gemini 3.1 Pro | Yes | Fast, generous free tier inside Google AI Pro at $19.99/mo | 1M tokens |
For high-stakes captions where a wrong number is fatal, Claude Opus 4.7 is the safer default: it is tuned to say “this value is not legible” rather than invent one. For batch captioning of a clean deck, any of the three works; pick the one you already pay for. (Free tiers: ChatGPT Free shows ads in the US as of February 2026 and has tight limits; Claude Free runs limited Sonnet 4.6.)
Copy-ready prompt
Paste the chart image, then this. Replace each [bracket] with your real values; keep the brackets out of the final text.
Write a 2-sentence takeaway for the chart image above.
Chart facts (use these numbers, do not read them off the image):
- Type: [bar / line / scatter / stacked]
- X-axis: [what + range]
- Y-axis: [what + units]
- Series: [list]
- Key numbers: [first, last, peak, dip, % change to quote]
Audience: [who reads this + the decision they own]
My hypothesis: [optional one line]
Output rules:
- Sentence 1: the single most decision-relevant pattern, in plain language, with one specific number from the facts above.
- Sentence 2: the "so what" — what this implies for the audience's decision.
- 25 to 45 words total. No axis labels, no chart-structure description, no vague verbs like "trends upward".
- If a number I need is not in the facts, ask for it instead of guessing.That last line is the one that saves you: it turns a confident hallucination into a quick question.
What good output looks like
Sentence 1 names the decision-relevant pattern with one concrete number. Sentence 2 closes the loop: what the reader should do, watch, or question. Total length 25 to 45 words; anything longer is the model hedging.
- Weak: “The chart shows revenue across four quarters, with Q4 being the highest bar.”
- Strong: “Revenue climbed 12% to $4.2M in Q4, but gross margin held at 38% all year, so the next lever is price, not volume.”
How to check the output
- Read it without the chart in front of you. If it only makes sense with the chart open, it is too thin.
- Confirm every number traces back to the facts you typed, not to something the model read off the image.
- Check that the “so what” ties to the audience’s actual decision, not a generic “worth monitoring.”
Caption the whole deck in one pass
The most common defect is captions that contradict each other across a deck. Solve it by captioning in one conversation: give the model the deck’s overall narrative first, then send charts one message at a time. It keeps the running story in context and stops chart 7 from undercutting chart 2. With a 1M-token context window on Opus 4.7 or Gemini 3.1 Pro, a 30-chart deck fits comfortably in a single thread.
FAQ
- Can I paste a screenshot directly? Yes. GPT-5.5, Opus 4.7, and Gemini 3.1 Pro all read chart images. Still type the key numbers as a fallback for numeric accuracy.
- Which model hallucinates numbers least? Opus 4.7 is tuned to report unreadable values instead of inventing them, which makes it the safer pick for high-stakes captions as of June 2026.
- How short should captions be? 25 to 45 words on a slide; under 25 on a dashboard tile where space is tight.
- Does every chart need a caption? Only load-bearing charts. A decorative or context chart does not need a takeaway.
- Does the free tier work for this? It can, with limits. ChatGPT Free (with US ads since February 2026) and Claude Free both read images but cap usage; for a full deck, a paid tier avoids mid-session throttling.
Related
- AI explains charts and tables
- AI writes dashboard takeaways
- AI writes weekly KPI summary
- Excel / Spreadsheet Analysis Prompts
Tags: #Data analysis #Workflow #Research