Matrixx0070 Bulk Extract
Fan the same extraction goal out across many URLs, collect the results, and hand back one clean table.
Pre-flight Check (REQUIRED)
Run the use-tinyfish pre-flight (CLI installed + authenticated). Stop with install/login instructions if either fails.
Workflow
1. Assemble the inputs
Three things are needed — gather what's missing:
- URL list — from the user's message, an uploaded file, or discovered via
tinyfish search queryper entity (e.g. one search per company name to find its pricing page) - Extraction goal — one JSON spec applied to every URL
- Output columns — the fields of that JSON, plus
urlandstatus
Confirm the goal on ONE sample URL before fanning out — run it, show the user the extracted row, adjust the spec if needed. This avoids burning a whole batch on a bad goal.
2. Choose the cheap path first
If the pages are static content, skip agent runs entirely:
tinyfish fetch content get --format markdown <url1> <url2> <url3> ...
Fetch accepts many URLs in one parallel call — extract the fields from the returned markdown yourself. Reserve batch agent runs for pages that are dynamic, interactive, or where fetch comes back empty.
3. Batch agent runs
Build a CSV with url,goal columns (same goal each row, JSON quotes escaped), then:
tinyfish agent batch run --input runs.csv
tinyfish agent batch get <batch_id> # poll until terminal
Poll at a sensible interval and report progress at milestones (e.g. "23/40 complete"), not on every poll. Collect each run's resultJson; note per-run status.
For small lists (≤5 URLs), parallel individual tinyfish agent run calls are fine — skip the batch machinery.
4. Handle failures
- Retry failed runs individually with
tinyfish agent run(once). - Rows that still fail stay in the output with
status: FAILEDand an empty data cell — never silently drop URLs. Report the failure count and list up front.
5. Build the deliverable
Normalize all results into one table: one row per URL, consistent columns, numbers as numbers, dates in ISO format, a status column (OK / FAILED / EMPTY).
If a spreadsheet-authoring skill is available in this environment, read it first and produce an .xlsx (frozen header row, sensible column widths, failures highlighted). Otherwise produce a clean .csv. Name the file after the task (e.g. competitor-pricing-2026-07-17.xlsx).
6. Deliver and summarize
Send the file with a short summary: rows extracted, failures, and one or two notable observations from the data (highest/lowest, obvious outliers). Offer next steps where natural — turn it into a dashboard (web-dashboard) or watch a subset for changes (web-watch).
$ARGUMENTS