Local Government Policy Analysis — Master Orchestrator

Before anything else: read the context file

This skill is jurisdiction-neutral. Every dataset identifier, field name,
office title, fiscal calendar and geographic unit it needs lives in
context.yml beside this file, copied from context.template.yml.

Read the filled context file before the first tool call. If it still
contains TODO values that the task depends on, say which ones and stop —
guessing a dataset ID or a field name produces a confident wrong answer with
no error to catch it.

Two conventions used throughout:

• Geographic unit — your jurisdiction's primary sub-area for analysis:
  neighborhood, ward, district, borough, community area. Named in the context
  file as geographic_unit_label, carried in data as {{geographic_unit_field}}.
• Service requests — your non-emergency request system, whatever it is
  called locally (311, FixIt, CRM, work orders). Datasets and fields are in the
  context file.

Scope: this skill assumes a US local government — 311-style service
requests, ZIP codes, UCR/NIBRS crime classification, US plain-language reading
conventions. Adopters elsewhere will need to adapt those, not merely fill slots.

---

Five-Phase Integrated Framework

Phase	Source Methodology	When to Use	Reference File
1. FRAME	Bloomberg Center for Public Innovation (JHU)	Problem is undefined or needs scoping	Problem_Framing_Skill.md
2. ANALYZE	J-PAL, MIT — Evidence-to-Policy	Running numbers, finding patterns, equity analysis	Analytical_Skill.md
3. COMMUNICATE	The GovLab (NYU) / InnovateUS	Writing memos, briefs, dashboards, community reports	Communication_Skill.md
4. BENCHMARK	Cross-jurisdiction comparison against the peer set in your context file	Comparing your jurisdiction to peers, learning from elsewhere	Benchmarking_Skill.md
5. PERFORM	Results for America / PerformanceStat (CitiStat)	Budget × staffing × service outcomes: cost-per-outcome, workload-per-FTE, efficiency trends	Performance_Management_Skill.md

Always read the relevant sub-skill file before beginning each phase.

---

Quick Decision Router

User Request
│
├─ "What data do we have on..." / "Help me define the problem"
│   → Read Problem_Framing_Skill.md → Run Phase 1
│
├─ "Analyze / run the numbers / what does the data show / is there an equity issue"
│   → Read Analytical_Skill.md → Run Phase 2
│
├─ "Write a memo / create a brief / make a dashboard / present these findings"
│   → Read Communication_Skill.md → Run Phase 3
│
├─ "Compare us to other cities, counties or towns / how do we rank / what works elsewhere"
│   → Read Benchmarking_Skill.md → Run Phase 4
│
├─ "Budget vs. performance / cost per outcome / workload per FTE / are we getting results / staffing efficiency / how much does it cost to / is the department understaffed / overtime analysis"
│   → Read Performance_Management_Skill.md → Run Phase 5
│
└─ "Full analysis / investigate / give me a recommendation / policy project"
    → Run all relevant phases in sequence

---

MCP Tool Reference

Primary jurisdiction

Your portal is named in the context file as primary_portal_url — cite it in
every published output. Its MCP server and four core calls are
primary_mcp_server, tool_search_datasets, tool_get_dataset_info,
tool_get_schema and tool_query. The pattern is the same on every platform:

{{tool_search_datasets}}    → Discover datasets by topic
{{tool_get_dataset_info}}   → Find resource IDs and metadata
{{tool_get_schema}}         → Get exact field names before querying
{{tool_query}}              → Retrieve records

Peer jurisdictions

Each peer's MCP server, platform and query syntax is listed under
comparators.peer_jurisdictions in the context file. Platforms differ — Socrata,
CKAN and ArcGIS take different call shapes and different field conventions —
so read the peer's entry before calling it.

⚠️ ALWAYS confirm field names via schema before querying any dataset in any
jurisdiction. Field names differ across jurisdictions, across datasets within
one jurisdiction, and across years within one dataset.

---

Standard MCP Sequence

1. Search datasets by topic          → find dataset IDs
2. Get dataset info                  → find queryable resource IDs
3. Get schema                        → confirm EXACT field names
4. Query                             → retrieve records

Step 3 is not optional. Skipping it is the most common cause of a query that
returns zero rows and gets read as a real finding.

---

System changes and schema drift

Service request systems get replaced, and field names rarely survive the
migration. Budget datasets roll their column names forward each cycle.
Department names are recorded inconsistently across datasets in almost every
jurisdiction.

Three rules follow:

1. Confirm the schema for each resource you touch, not once per session.
2. When a system was replaced mid-period, query each era separately and
   reconcile. Do not join across a migration and report a single trend line
   without saying so.
3. Verify the exact department string before filtering on it. A department
   filter that silently matches nothing looks identical to a department with no
   cases.

Any known migration dates and field mappings for your jurisdiction belong in
the context file under service_request_resource_ids.

---

Cross-Phase Quality Standards

Rigor (J-PAL)

Every claim is grounded in data or clearly labeled as interpretation.
Confidence level stated. Limitations named, not buried.

Human-Centeredness (Bloomberg)

Problem framed around people's lived experience. Recommendations are
implementable by the staff you actually have — check analytics_function and
departments in the context file before recommending work nobody is resourced
to do.

Inclusivity (GovLab)

Equity lens applied. Plain-language versions exist. Feedback mechanisms
included, with a working contact route from the context file.

Transparency

Data sources cited with IDs and portal URL. Methodology reproducible. Findings
shareable as open knowledge.

---

Supporting Files in This Skill Set

File	Purpose
context.template.yml	The blank template. Copy to context.yml and fill it in first — every jurisdiction-specific value the skill needs
Problem_Framing_Skill.md	Bloomberg methodology: scope, stakeholders, assumptions
Analytical_Skill.md	J-PAL methodology: descriptive → diagnostic → equity
Communication_Skill.md	GovLab/InnovateUS: memos, briefs, dashboards, engagement
Benchmarking_Skill.md	Cross-jurisdiction comparison; includes Performance Management Benchmarking module
TEMPLATES.md	Fill-in-the-blank templates for 6 output types
CHECKLISTS.md	Pre-flight and review checklists for all phases
PROMPTS.md	Example prompts organized by phase and complexity
REFERENCE.md	How to build and maintain your dataset directory
Performance_Management_Skill.md	Results for America / PerformanceStat: budget × staffing × outcomes efficiency analysis
EXAMPLE-service-equity.md	Complete worked example: service request response equity analysis

---

When Creating Documents

• Word docs (.docx): also read the docx skill
• Presentations (.pptx): also read the pptx skill
• Spreadsheets (.xlsx): also read the xlsx skill
• Dashboards (React/HTML): also read the frontend design skill available in
  your environment

Skill paths differ by environment; locate them rather than assuming a fixed path.