| name | save_correction |
| description | Use this skill when the user identifies that knowledge in DataHub is wrong, incomplete, or missing — whether that's a glossary term definition, a domain or data product description, a dataset or column description, an existing document that needs updating, or a concept that needs a new reference document. Also triggered by "fix this description", "update the metadata", "correct the definition", "update the glossary", "fix the domain description", or "write that down".
|
| license | MIT |
| compatibility | Requires DataHub write-back enabled in Settings → Connections |
| allowed-tools | search, get_entities, list_schema_fields, search_documents, save_correction |
| metadata | {"author":"analytics-agent","version":"2.0"} |
save_correction
Overview
This skill writes correct knowledge back to DataHub. It handles three cases:
| Mode | When to use | Key inputs |
|---|
| Entity/field description | A glossary term, domain, data product, dataset, or column description is wrong, incomplete, or missing | entity_urn + corrected_description |
| Update existing doc | An existing DataHub document has wrong or outdated content | doc_urn + doc_title + doc_body |
| Create new doc | No document exists for this concept; needs to be written from scratch | doc_title + doc_body + parent_doc_urn |
Always confirm with the user before writing anything.
Mode 1 — Entity / field description
Supports any DataHub entity type that has a description field:
| Entity type | URN format | How to find the URN |
|---|
| Glossary term | urn:li:glossaryTerm:<id> | search(filter="entity_type = glossaryTerm") |
| Domain | urn:li:domain:<id> | search(filter="entity_type = domain") |
| Data product | urn:li:dataProduct:<id> | search(filter="entity_type = dataProduct") |
| Dataset | urn:li:dataset:(...) | search or get_entities |
| Column (field) | same dataset URN + field_path | list_schema_fields on the dataset |
Step 1 — Confirm the entity URN
The entity_urn must come from a prior search, get_entities, or
search_business_context result. Never construct a URN from scratch.
Step 2 — Fetch the current description
Glossary term / domain / data product: Call get_entities([entity_urn]) and
extract description or properties.description.
Dataset-level: Call get_entities([entity_urn]) and extract description.
Field-level: Call list_schema_fields on the dataset URN; note the field's
current description and exact fieldPath value.
If no description exists, note that one will be added.
Step 3 — Show the user what will change
Entity: <dataset name> (<urn>)
Field: <field name> (if applicable)
Current description: <current text, or "(none)">
Proposed correction: <new description>
Shall I apply this correction?
Do not call save_correction until the user confirms.
Step 4 — Choose operation
| Situation | operation |
|---|
| Replacing wrong/outdated text | "replace" (default) |
| Adding a clarifying note | "append" |
Step 5 — Call save_correction
Pass: entity_urn, corrected_description, field_path (if field-level), operation.
Mode 2 — Update existing doc
Step 1 — Retrieve the document
Use the doc_urn from a prior search_documents result. If you don't have it,
call search_documents with the document title or topic to find it.
Step 2 — Show the diff
Present the existing document title and a summary of what will change. Ask for
confirmation before writing.
Step 3 — Call save_correction
Pass: doc_urn, doc_title (may be unchanged), doc_body (full updated markdown body).
Mode 3 — Create new doc
Step 1 — Find the right parent
Before creating, call search_documents to find the most relevant existing folder
or document that should be the parent. Examples:
- Fixing docs about "Revenue Metrics" → find the existing "Revenue" or "Finance" folder
- Adding a definition for a specific dataset → find the dataset's documentation folder
- General concept with no clear home → find the top-level "Knowledge Base" or "Data Dictionary" folder
If no suitable parent exists, use the org's top-level shared folder (search for
"Shared" or "Knowledge Base"). As a last resort, omit parent_doc_urn and
the document will be created at the root level.
Step 2 — Draft the document
Write a well-structured markdown document. Use headings, bullet points, and
code blocks where appropriate. For concept/metric definitions, use this template:
## Definition
<clear 1–2 sentence definition>
## How it's calculated
<formula or logic>
## Source tables
- <table name> — <what it contributes>
## Common pitfalls
- <gotcha 1>
- <gotcha 2>
Adapt freely — the template is a starting point, not a requirement.
Step 3 — Confirm with the user
Show: proposed title, target parent folder (name + URN), and the full document body.
Ask for confirmation before writing.
Step 4 — Call save_correction
Pass: doc_title, doc_body, parent_doc_urn (from Step 1),
related_entity_urns (dataset URNs referenced in the doc, if any).
After writing (all modes)
Report back:
- Whether it succeeded
- The URN of the updated/created entity or document
- For docs: the title and parent folder so the user can find it in DataHub
