Screenshot to Structured Data

Convert image content into a checked, reusable deliverable: a table, task list, or field mapping. Treat screenshot understanding as an extraction task, not a prose-summary task.

Workflow

1. Inspect every supplied image before proposing a schema. Record its filename or a short source label so the output can be traced back to the image.
2. Select one output mode:
   • table for records such as orders, leads, products, expenses, or survey responses.
   • tasks for chat screenshots, meeting captures, and requirement discussions.
   • field-map for forms, admin pages, and data-entry screens. Read references/output-contracts.md before preparing the export payload.
3. Confirm the minimum meaningful columns. Infer an obvious schema without asking; ask one concise question only when two materially different schemas are equally plausible. When no records are visible, still choose the requested or default columns and export an empty result.
4. Extract source facts faithfully. Preserve identifiers, dates, amounts, statuses, and quantities exactly as shown. Include personal contact details only when needed for the user's requested result; otherwise redact or omit them. Use an empty value or Unknown for unreadable information; never manufacture a value.
5. Normalize only when it improves the requested deliverable. Keep raw_* values when normalizing dates, currencies, names, or status labels could lose source meaning.
6. Add source to every extracted record. Use a top-level source only when the same source applies to every record. Put ambiguous, partially visible, or conflicting content in review_needed rather than silently guessing.
7. Build a JSON payload matching the selected contract and export it with:

$script = "<installed-skill-root>/scripts/export_structured_data.py"
python $script payload.json output.xlsx

Locate <installed-skill-root> from the directory containing this SKILL.md; do not assume a project-relative scripts/ directory. Use --format csv, --format json, or --format md when the requested deliverable needs a different format. CSV exports with review items also create a *-review-needed.csv sidecar.

Quality Gate

Before delivering the result:

• Reconcile the number of extracted records with the visible source items.
• Recheck high-impact cells: IDs, amounts, dates, contact details, totals, and owners.
• Ensure every task has a concrete action; leave owner or due date blank when absent instead of inventing one.
• State the extraction mode, every output path, and any items requiring review.

Do not claim OCR-grade certainty for content that is blurred, cropped, obscured, or too small to inspect.