pbi-prompt-builder

name	pbi-prompt-builder
description	Interactive Power BI prompt builder: discovers semantic models, explores schema, drafts DAX queries from natural-language questions, validates them live, and outputs a ready-to-use pbi-*.prompt.md file. Triggers: build PBI prompt, create Power BI report prompt, scaffold PBI workflow, map Power BI report, generate PBI prompt, new PBI query, Power BI prompt template, what data can I pull.
argument-hint	Describe the questions you want to answer, or provide a semantic model name/ID to explore

Purpose

Guides users through building a reusable Power BI prompt file (pbi-*.prompt.md) by interactively discovering what data is available, drafting DAX queries for their questions, validating them against live data, and assembling the result into a prompt that follows project conventions.

When to Use

• User wants to create a new PBI-backed workflow (incentive tracking, gap analysis, pipeline scoring, consumption review)
• User knows the questions but not the DAX or schema
• User wants to map an existing Power BI report into a repeatable agent prompt
• Manager wants to customize a PBI prompt for their team

Freedom Level

High — Discovery and question refinement require judgment. DAX generation uses powerbi-remote:GenerateQuery. Output structure is Low freedom (must follow template exactly).

Runtime Contract

• Tools: powerbi-remote:DiscoverArtifacts, powerbi-remote:GetSemanticModelSchema, powerbi-remote:GetReportMetadata, powerbi-remote:GenerateQuery, powerbi-remote:ExecuteQuery
• Auth convention: Follow powerbi-mcp.instructions.md § Auth Pre-Check Pattern
• Output convention: Follow powerbi-mcp.instructions.md § Prompt Template Convention (pbi-*.prompt.md naming)

Flow

Phase 1 — Intent Gathering

Ask the user (skip items they've already provided):

1. What questions do you want to answer? — Collect 1–5 natural-language questions. Examples:

   • "What is my gap to target for account X?"
   • "Which opportunities in my pipeline have the highest conversion likelihood?"
   • "Show me accounts trending below consumption threshold"
   • "How is my team tracking against incentive Y?"

2. Which Power BI report or semantic model? — Options:

   • User provides a dataset ID (GUID) → use directly
   • User provides a report/model name → call powerbi-remote:DiscoverArtifacts to resolve
   • User doesn't know → call powerbi-remote:DiscoverArtifacts with keywords from their questions, present matches

3. Account scope — Where do TPIDs/account identifiers come from?

   • A vault file (e.g., .docs/AccountReference.md)
   • CRM (msx:get_my_active_opportunities)
   • User will provide inline
   • Not applicable (model-wide query)

4. Business rules or reference docs — Any program rules, incentive definitions, or threshold documents to embed? (Optional — stored in .github/documents/)

Phase 2 — Schema Discovery

1. Auth pre-check — per powerbi-mcp.instructions.md:

   EVALUATE TOPN(1, 'Dim_Calendar')
   

   If this fails → stop and show the auth recovery message. Do not proceed.

2. Get schema — call powerbi-remote:GetSemanticModelSchema({ artifactId }).

3. Inspect report (if user referenced a report) — call powerbi-remote:GetReportMetadata({ reportObjectId }) to understand how the model is used in practice: which tables, filters, and measures the report author intended.

4. Present a plain-language summary of what's in the model:

   • Key tables and what they represent
   • Available measures (pre-built calculations)
   • Important filter dimensions (calendar, account, segment, etc.)
   • Any custom instructions or verified answers from the model author

5. Map questions to schema — for each user question, identify:

   • Which tables/columns are needed
   • Which measures apply
   • What filters are required
   • Flag any questions that can't be answered by this model

Present the mapping and ask: "Does this look right? Any questions to add, change, or drop?"

Phase 3 — DAX Generation & Validation

For each confirmed question:

1. Generate DAX — call powerbi-remote:GenerateQuery with:

   • artifactId: the semantic model ID
   • userInput: the user's natural-language question
   • schemaSelection: the table/column/measure mapping from Phase 2
   • valueSearchTerms: any specific account names, product names, etc.

2. Test-run — call powerbi-remote:ExecuteQuery with the generated DAX, limited to a small result set (maxRows: 10) to validate it returns data.

3. Show results — present the sample data to the user. Ask:

   • "Is this the data you expected?"
   • "Should we add/remove columns?"
   • "Any filters to adjust?"

4. Iterate — if the user wants changes, re-generate with updated userInput or schemaSelection. Use chatHistory parameter to carry prior DAX context forward.

5. Parameterize — replace hardcoded values (specific TPIDs, dates) with placeholders that the prompt workflow will fill dynamically:

   • TPID lists → {"<TPID1>", "<TPID2>", ...} (injected from account roster)
   • Dates → RelativeFM offsets or computed values
   • Thresholds → Configuration table variables

Phase 4 — Prompt Assembly

Generate the .prompt.md file following this structure:

---
description: "<one-line description of what this prompt does>"
---

# <Title>

<One-sentence instruction to the agent>

## Reference

<Link to any .github/documents/ files with program rules. Omit if none.>

## Key Rules Summary

<Business rules extracted from reference docs or user input. Omit if not applicable.>

## Configuration

> **Managers**: Fork this file and update these values.

| Setting | Value | Notes |
|---|---|---|
| **Semantic Model ID** | `<GUID>` | <model name> |
| **Account Roster** | `<path>` | <description> |
| <additional settings> | | |

## Workflow

### Step 0 — Power BI Auth Pre-Check
<standard auth check — always include>

### Step N — <Step Title>
<workflow step with DAX query, logic, or cross-medium retrieval>

### Step N+1 — Report
<output format — table layout and status values>

Phase 5 — Write & Confirm

1. Propose filename — pbi-<kebab-case-name>.prompt.md in .github/prompts/.
2. Write the file — create it in the workspace.
3. Summary — show what was created and how to use it:
   • "Run this prompt by asking: <trigger phrase>"
   • "To customize: edit the Configuration table in the file"
   • "To change the data source: swap the Semantic Model ID"

Decision Logic

Situation	Action
Model has no relevant tables for a question	Drop the question; explain what the model can answer
GenerateQuery produces invalid DAX	Retry with broader schemaSelection; if still fails, hand-write DAX using schema
ExecuteQuery returns empty results	Check filters — may be scoping too narrowly. Show the user and ask for guidance
User wants cross-medium data (PBI + CRM)	Add CRM steps to the workflow (account roster, opp state); reference shared-patterns skill
User wants to use vault for account scoping	Add vault read step; reference obsidian-vault.instructions.md
Multiple semantic models needed	Create one prompt per model, or a multi-step prompt with separate queries per model

Output Schema

• prompt_file_path: path to the created .prompt.md file
• questions_mapped: list of questions → DAX query pairings
• questions_dropped: any questions the model couldn't answer (with explanation)
• next_action: "Prompt created. Run it with: <trigger phrase>. To refine, edit the Configuration table or re-run this skill."