Imported from Stratio/genai-agents (skills/manage-critical-data-elements/SKILL.md). Install upstream with npx skills add Stratio/genai-agents --skill manage-critical-data-elements. Copyright stays with the author.

Skill: Manage Critical Data Elements

Workflow for consulting and defining Critical Data Elements (CDEs) in a governed domain. CDEs are the data assets — tables or specific columns — considered most important for the business. They receive prioritized quality treatment: assessments focus on them first, and gaps in CDE assets are escalated to a higher priority level.

CRITICAL APPROVAL PAUSE

NEVER call set_critical_data_elements without explicit user confirmation.

HTTP 409 responses during tagging mean the asset was already tagged as a CDE. This is NOT an error — count these as "already tagged" and do not treat them as failures in results or summaries.

1. Scope Determination

1.1 Validate the domain

If $ARGUMENTS provides a domain name, search with search_domains($ARGUMENTS). If not found, retry with search_domains($ARGUMENTS, refresh=true). If still not found, or no argument was provided, list domains with list_domains() and ask the user with options following the user question convention.

CRITICAL rule: the domain_name used in all MCP calls must be exactly the value returned by search_domains or list_domains. NEVER translate, interpret, paraphrase, or infer it.

1.2 Determine flow

If the user's request already makes the intent clear, proceed directly:

• "Show me the CDEs" / "What are the critical data elements of [domain]?" → Flow A
• "Recommend CDEs" / "Define CDEs" (no specific asset mentioned) → Flow B (ask method in B.2)
• "Tag [specific table/column] as critical" (user names exact tables or columns) → Flow B, skip B.2, go directly to B1 with those assets pre-filled. Do NOT ask the user to choose a method — they already specified what to tag.
  • User named specific columns → pre-filled assets go in columns_by_table, NOT in critical_tables.
  • User named whole tables (no column list mentioned) → pre-filled assets go in critical_tables.

Otherwise, ask the user with options following the user question convention:

1. View current CDEs — consult what is currently marked as critical in this domain
2. Define/update CDEs — tag tables or columns as Critical Data Elements

Flow A: View Current CDEs

A.1 Retrieve and present current CDEs

Call get_critical_data_elements(domain_name=domain_name).

Present results with clear structure:

## Critical Data Elements — [domain_name]

### Fully critical tables (all columns are CDEs)
| Table |
|-------|
| account |
| district |

### Tables with specific CDE columns
| Table | CDE Columns |
|-------|-------------|
| card | card_id, disp_id |
| transaction | account_id, bank, date, trans_id |

If both critical_tables and columns_by_table are empty or absent: "No Critical Data Elements are currently defined for this domain."

A.2 Optional: quick quality coverage check

After presenting CDEs, offer the user a quick check of quality coverage over the CDE assets. If accepted:

1. Call get_tables_quality_details(domain_name, [all CDE tables]) — tables from critical_tables plus tables with entries in columns_by_table
2. Present a summary table:

| Table | CDE type | Rules defined | Status |
|-------|----------|---------------|--------|
| account | Full table | 4 | OK |
| card | Specific columns | 1 | Warning |
| transaction | Specific columns | 0 | No rules |

3. If any CDE table has no rules defined: highlight as a quality gap to address.

A.3 Follow-up question

Ask the user with options following the user question convention:

• Define or update CDEs for this domain (proceed to Flow B)
• Assess quality coverage focused on CDEs (load assess-quality)
• Finish

Flow B: Define/Update CDEs

B.1 Gather baseline in parallel

Before recommending or collecting user input, fetch the current state:

Parallel:
  A. list_domain_tables(domain_name)
  B. get_critical_data_elements(domain_name=domain_name)   ← baseline: current CDEs

B.2 Determine definition method

Ask the user with options following the user question convention:

1. Manual specification — I will indicate directly which tables/columns to tag as CDEs
2. Agent recommendation — Analyze the domain semantics and data profile, then recommend CDEs for my confirmation
3. Combined — Agent recommends, then I adjust before confirming

Flow B1: Manual Specification

Present the list of available domain tables (from list_domain_tables) to help the user choose. If the user wants to inspect columns for a specific table, call get_table_columns_details(domain_name, table) on request.

Collect from the user:

• Which entire tables to tag as critical (all columns in those tables become CDEs)
• Which specific columns in which tables to tag as CDEs

Validate that all mentioned tables and columns exist in the domain before proceeding. Compare the user's specified assets against the baseline from B.1:

• Assets not in the baseline → include in the tagging payload (truly new)
• Assets already in the baseline → exclude from the payload; note them as "Already tagged" in the plan

Payload rule: the tagging payload must contain ONLY the assets that are not in the baseline. If the user requested N assets but all N are already tagged, the payload is empty — skip the API call entirely and present the result as "all requested assets were already tagged as CDEs".

Proceed to B.4 Approval Pause.

Flow B2: Agent Recommendation

B2.1 Gather information in parallel

Parallel:
  A. get_tables_details(domain_name, [all_tables])
  B. get_table_columns_details(domain_name, table)          [for each table — in parallel]
  C. get_tables_quality_details(domain_name, [all_tables])  ← existing rules as recommendation signal

Then profile the data for each table:

For each table (in parallel):
  generate_sql("get all fields from table [table] without filters", domain_name)
  → profile_data(query=[sql])

B2.2 Recommendation criteria

Table-level (tag entire table as CDE):

• Master or reference entities (accounts, customers, products, locations, categories)
• Tables described in the domain as foundational or central to business processes
• Tables with many foreign key relationships pointing to them from other tables
• Tables explicitly mentioned in documented business rules as authoritative sources of truth

Column-level (tag specific columns within a table):

• Primary keys and business identifiers — completeness + uniqueness are critical by definition
• Mandatory foreign keys that drive referential integrity across the domain
• Columns defined in the domain glossary with business terms or formal definitions
• Core financial or operational metric columns (amounts, counts, rates central to reporting)
• Date columns that determine data freshness, SLA compliance, or regulatory reporting windows
• Status or state columns that control business process flows or approvals
• Columns already covered by multiple quality rules — prior investment signals importance

EDA signals that reinforce a recommendation:

• Existing nulls in a column that should be non-null → immediate evidence of risk
• Duplicates in a primary key candidate → critical integrity concern
• Columns with several quality rules already defined → someone already considered them critical

B2.3 Present recommendation

## Recommended Critical Data Elements — [domain_name]

**Basis**: semantic analysis of domain description, table context, column definitions,
domain terminology, EDA results, and existing quality rules.

### Recommended fully critical tables (all columns become CDEs)

| Table | Reason |
|-------|--------|
| account | Master entity — all columns are identifiers, balances, or process-critical fields |
| district | Reference table — provides geographic master data used across 3+ tables |

### Recommended tables with specific CDE columns

| Table | Recommended columns | Reason |
|-------|---------------------|--------|
| card | card_id, disp_id | Primary key + mandatory FK — core integrity identifiers |
| transaction | account_id, trans_id, date, amount | FK, business ID, freshness date, core financial metric |

### Tables not recommended as CDEs

| Table | Reason |
|-------|--------|
| order | Secondary operational table; no master/reference role identified |

Inform the user they can accept, modify, or reject individual items before confirming.

Flow B3: Combined

Execute Flow B2 (agent recommendation) first, then allow the user to:

• Add tables or columns not included in the recommendation
• Remove tables or columns from the recommendation
• Promote a column-level CDE to a full-table CDE (or demote)
• Accept the recommendation as-is

Collect the final agreed list, then proceed to the approval pause.

B.4 PAUSE: Present Plan and Wait for Approval

Before calling set_critical_data_elements, present the complete tagging plan:

• Assets to tag as critical = only assets NOT present in the baseline (truly new tagging operations).
• Already tagged = assets from the user's specification that were already in the baseline — shown for transparency, excluded from the API call.

## CDE Tagging Plan — [domain_name]

**Domain**: [domain_name]
**Source**: [Manual specification | Agent recommendation | Combined]

### Assets to tag as critical (new — not yet in the baseline)

**Full critical tables** (all columns will be CDEs):
- account
- district
- loan

**Tables with specific CDE columns**:
| Table | Columns to tag |
|-------|----------------|
| card | card_id, disp_id |
| transaction | account_id, bank, date, trans_id |

**Already tagged** (from current baseline — will be counted as retained, not errors):
- [list or "None"]

---

Shall I proceed with tagging these assets as Critical Data Elements?

Valid approval signals: "yes", "proceed", "go ahead", "ok", "approved", "confirmed", "do it", or equivalents in the user's language.

Rejection or modification signals: "no", "cancel", "remove X", "add Y", "change". If the user modifies: update the plan, present again, and wait for confirmation.

B.5 Execution

Only after explicit approval, call set_critical_data_elements with only the assets that are new (not present in the baseline from B.1).

CRITICAL — payload must contain ONLY new assets:

• Cross-check the payload against the baseline obtained in B.1 (get_critical_data_elements result).
• Remove from the payload ANY asset (table or column) that already appears in the baseline.
• DO NOT include currently-tagged CDEs. The API is additive, not a full replace: re-sending an existing CDE causes a 409 error and is never necessary.
• If the user's specification matches only 1 new asset, the payload contains exactly 1 asset. If it matches 0 new assets, skip the API call entirely.

Column vs table distinction: if the user named specific columns (not whole tables), those columns MUST appear in columns_by_table — never in critical_tables. Using critical_tables for a column-level request marks the entire table as CDE.

The call takes a single collection object (not three separate parameters):

set_critical_data_elements(collection={
  "domain_name": "domain_name",
  "critical_tables": ["account", "district", "loan"],
  "columns_by_table": {
    "card": ["card_id", "disp_id"],
    "transaction": ["account_id", "bank", "date", "trans_id"]
  }
})

B.6 Present results

Parse the response and present a result table:

## CDE Tagging Result — [domain_name]

| Asset | Result |
|-------|--------|
| account (full table) | Tagged successfully |
| district (full table) | Already tagged — no change needed |
| card.card_id | Tagged successfully |
| card.disp_id | Tagged successfully |
| transaction.account_id | Already tagged — no change needed |
| transaction.bank | Tagged successfully |

**Summary**:
- Successfully tagged: N
- Already tagged (retained): M
- Failed: X

• HTTP 409 → show as "Already tagged — no change needed" and count as retained, not as failure
• Other errors → show as "Failed" with a brief reason note below the table

2. Follow-up Question

Ask the user with options following the user question convention:

• Assess quality coverage focused on CDEs (load assess-quality)
• View the updated list of CDEs for this domain
• Define CDEs for another domain
• Finish