Skip to content
Skills Library
Documents & office · Automation

document-understanding

/document-understanding

Reads fields and tables out of documents (invoices, receipts, purchase orders, statements, IDs, payslips, forms, any PDF or scan) with a Document Understanding reader, by conversation with a non-technical person: set up or reuse a reader, confirm its fields in plain words, try it on one sample, then build the flow with the Document Understanding nodes and a human check for uncertain documents. Use when the person wants values read from documents.

Licence
MIT
Last repository change
27.10.2026

Assessment and security

100/ 100Class: good
Format25 / 25
Frontmatter with name and description, valid name, description up to 1,024 characters, name matching the folder.
Discoverability25 / 25
Description explains when to use the skill, has a reasonable length and uses third person.
Content35 / 35
At least 150 words and at most 500 lines, headings, step lists, examples or code.
Hygiene15 / 15
No placeholders such as TODO, no Windows paths, with a licence.
Riskno signals detected in the scan
Files in this package
2
Scripts
0
HTML comments
0
URLs in SKILL.md
0
Deprecation

No signals detected. Repository changed on 27.10.2026, without references to retired models or APIs.

Licence

MIT, open: use and modify while retaining author attribution.

What it does and when to use it

Reads fields and tables out of documents (invoices, receipts, purchase orders, statements, IDs, payslips, forms, any PDF or scan) with a Document Understanding reader, by conversation with a non-technical person: set up or reuse a reader, confirm its fields in plain words, try it on one sample, then build the flow with the Document Understanding nodes and a human check for uncertain documents. Use when the person wants values read from documents.

Original author description

Reads fields and tables out of documents (invoices, receipts, purchase orders, statements, IDs, payslips, forms, any PDF or scan) with a Document Understanding reader, by conversation with a non-technical person: set up or reuse a reader, confirm its fields in plain words, try it on one sample, then build the flow with the Document Understanding nodes and a human check for uncertain documents. Use when the person wants values read from documents.

How it works

  1. Words
  2. The conversation (one question per turn)
  3. Making the reader better
  4. The flow
Original SKILL.md headings · 56 lines, 978 words
SKILL.md preview
---
name: document-understanding
description: "Reads fields and tables out of documents (invoices, receipts, purchase orders, statements, IDs, payslips, forms, any PDF or scan) with a Document Understanding reader, by conversation with a non-technical person: set up or reuse a reader, confirm its fields in plain words, try it on one sample, then build the flow with the Document Understanding nodes and a human check for uncertain documents. Use when the person wants values read from documents."
triggers: [invoice, receipt, payslip, purchase order, delivery note, utility bill, bank statements, scanned, document understanding, /\bocr\b/, /\b(read|extract|pull|capture|scan|parse|digiti[sz]e)\b[^.!?]*\b(pdfs?|documents?|forms?|statements?|bills?|scans?)\b/, /\bfrom (the|these|those|my|our|each|every) (pdfs?|scans?|documents?|forms?)\b/]
---

# Reading documents

A **document reader** (a Document Understanding project) knows what kind of documents to expect and which values to pull out of them. The flow sends each file to the reader and gets the values back, plus a yes/no for "a person should check this". You set the reader up with the `du_*` tools (Robomotion's Build view has them; elsewhere, create the reader in Robomotion's Documents section and use its id), then write the flow.

## Words

The person never hears: project, taxonomy, schema, template, field id, confidence, job, version, staging, F1, JSON, selector, or any id. Say **reader**, **the values it found**, **I'm not sure about**, **someone checks it**. Show values as a short list, as printed on the document. Ids in tool results are for `main.ts` only.

## The conversation (one question per turn)

1. **Where do the files come from?** Ask only if not said: a folder on this computer, an email inbox, a download. You need a real path to at least one sample; an attached PDF is not on disk, so ask for the folder instead.
2. **Reuse first.** `du_list_projects`. If a reader already reads this kind of document, use it and say so in one line.
3. **Set one up.** Pick the template yourself (`du_list_templates` if unsure; never show the list). With sample files: `du_create_project_from_samples` (1 to 3 samples). Without: `du_create_project` with the template, or with `fields` when nothing fits.
4. **Confirm the fields in plain words.** Name what it will read in one sentence ("It will read the invoice number, date, supplier, total and the line items."). Then ask at most one question per turn, only where the meaning is unclear: "Is 'Total' the amount to pay, including tax?", "Do you also need the PO number?" Apply answers with `du_update_schema`; a `description` saying where the value is on the page helps the reader.

   **Values that are not printed as they are** (a one-line summary, a category, a risk level, an end date that is only the start date plus "6 months"): set `work_it_out: true` on that field and say in its `description` how to work it out ("The start date plus the contract's term"). Tell the person in plain words: "I'll work out the end date from the start date and how long the contract runs." Guidance that holds for every document (the language to answer in, what counts as high risk) goes in `instructions`, not in each field.
5. **Try it on one sample.** `du_process`. Show what it found:
   > From *invoice-0412.pdf*: Invoice number INV-0412, Date 14 March 2026, Total 1,200.00 EUR. I'm not sure about the due date, so invoices like this one will go to a person to check.

   Ask "Does that look right?" If a value is wrong, fix the field's `description` (or type) with `du_update_schema` and try again, at most twice, then move on.
6. **Ask what happens next,** only what is still unknown: where the values go (an Excel file, a sheet, a folder of CSVs), and whether the robot should wait for a person to check uncertain documents or leave them for later.

## Making the reader better

When the person asks how well it reads, or a value keeps coming back wrong, run this loop, **at most two rounds**:

1. `du_review_status`. Measuring needs documents a person has checked and set aside for measuring. If there are none, say so plainly ("To tell how well it reads, someone needs to check a few documents first") and stop.
2. `du_measure`. Tell the person the result in words: "It reads 8 of 10 values right. It gets the end date right only about half the time."
3. Reword what is weakest: the field's `description`, or the shared `instructions`, with `du_update_schema` and `as_staging: true`, so the current reader keeps working while the new wording is tried.
4. `du_measure` with the new version the update reported, then `du_compare_versions`.
5. Better: `du_promote` it and say "The new wording reads the end date right more often, so I've switched to it." Worse or the same: leave things as they were and say so in one line.

If `du_process` says no computer picked the document up, say plainly that document reading is not running on any of their computers yet, and carry on building the flow: it will work once it is.

## The flow

Follow `creating-flow` for everything else (hex ids, `validate_flow`, `save_flow`). Look up each node's exact properties (`get_node_schema`, or `robomotion describe node <type>`), and pin the package with `f.addDependency('Robomotion.DocumentUnderstanding', '<a published version from robomotion describe package>')`.

| Node | Use |
|---|---|
| `Robomotion.DocumentUnderstanding.ProcessDocument` | `inFilePath`, `inProjectId` (the reader's id, `Custom('...')`). `optCreateReview: 'yes'` opens a check for a person when unsure (the default `'no'` never does). `optTag: 'staging'` reads with the wording being tried instead of the live one. Outputs `msg.result`, `msg.fields` (value per field id; a field inside a group has a dotted id, read it as `msg.fields['contract_info.end_date']`), `msg.needsReview`, `msg.reviewTaskId`. |
| `Robomotion.DocumentUnderstanding.WaitForReview` | `inTaskId: Message('reviewTaskId')`. Holds the flow until the person has checked it (`optTimeout`, default one hour; `optContinueOnTimeout: 'continue'` to go on instead of failing). Its output is `msg.validatedResult`; set `outResult: Message('result')` so both paths carry `msg.result`. |
| `Robomotion.DocumentUnderstanding.Export` | `inResult: Message('result')` gives `msg.fieldsTable` (`optFieldsLayout: 'columns'` for one row per document) and `msg.lineItemsTable` (`optTableID` picks the table). Both are `{columns, rows}`: write them with the nodes in `creating-flow`'s `./docs/patterns/data-tables.md`. |
| `CreateReviewTask`, `GetReviewResult`, `Digitize`, `Classify`, `Extract` | Rarely needed: a check sent without waiting, polling a check later, or the steps one by one. |

Shape, inside the usual loop over the files:

```
ProcessDocument -> Function "Needs a check?" (outputs: 2, msg.needsReview)
  port 0 (yes) -> WaitForReview -> Export
  port 1 (no)  -------------------> Export -> write the row (Excel / CSV / sheet)
```

A document the person rejects comes back with `msg.outcome === 'rejected'`: skip its row and log why.
Files in this package (2)
  • SKILL.md 7 KB
  • config.json 1 KB

Versions

The database has no other skills named document-understanding.

Database snapshot: 13.11.2026 · package version fca061202f23 (SHA-256 digest of the entire package)

Downloads in the last 30 days: 0 · ranking

Report a problem with this skill: harmful content, incorrect descriptions or an author’s removal request.

Similar skills

  • 90veryfi-documents-ai
    Real-time OCR and data extraction API by Veryfi (https://veryfi.com). Extract structured data from receipts, invoices, bank statements, W-9s, purchase orders, bills of lading, and any other document. Use when you need to OCR documents, extract fields, parse receipts/invoices,…
  • 95parse-integration
    Extract structured data from documents (invoices, receipts, purchase orders, bills of lading, bank statements, contracts, forms, ID documents, scans, photos of paper) with Parse by Conversion Tools, and build a production integration against its API. Covers getting an API key,…
  • 100azure-ai-document-intelligence-ts
    Extract text, tables, and structured data from documents using Azure Document Intelligence (@azure-rest/ai-document-intelligence). Use when processing invoices, receipts, IDs, forms, or building cu...
  • 100azure-ai-document-intelligence-ts
    Extract text, tables, and structured data from documents using Azure Document Intelligence (@azure-rest/ai-document-intelligence). Use when processing invoices, receipts, IDs, forms, or building custom document models.
  • 100data-extractor
    Extract structured data from any document format using unified document processing. Use when a user asks to extract data from a document, parse a PDF, pull structured data from files, convert documents to JSON or CSV, extract fields from invoices or forms, or scrape data from documents.
  • 100extract-document-data
    Extract structured, grounded fields from documents — values cite their page, missing values abstain instead of hallucinating. Use for parsing invoices, payslips, statements, contracts.

Feedback and reports

Your message goes to Bartosz Sokoliński. An email address is needed only for a reply. See the privacy policy.