Labs / Chapter 7
Chapter 7 makes data engineering a core act of prompt cartography. In this lab, you will use natural language to search for open datasets, evaluate their fitness, curate a shortlist, harmonize schemas, and build a small task-specific dataset that can support a defensible map.
Most maps inherit their truth claims from data long before design begins. This lab treats dataset discovery, reliability assessment, cleaning, and merging as design work. Your final product is not a polished map; it is a traceable task-specific dataset and the pipeline record that explains why it deserves to be mapped.
Choose a mission that requires more than one dataset. The final merged table should have 12-30 rows so the exercise stays manageable.
Map mission brief Topic: Place or study area: Audience: Decision or understanding the map should support: Spatial unit for the final dataset, if known: Minimum time period: Required fields or variables: Sensitive data or ethical guardrails: What the final map must not imply:
Start by showing yourself how much the wording matters. Use a normal web search or data portal search first, then use a structured LLM prompt.
Keyword-style search phrase: [Example: heat risk data Phoenix] Semantic dataset search prompt: Locate open datasets for [PLACE] that can support a map about [PURPOSE] for [AUDIENCE]. I need datasets that can be combined into a task-specific table at [GEOGRAPHY OR UNIT]. Return candidate datasets with: - dataset name - source organization - URL - format or access method - geometry or spatial fields - temporal coverage - key attributes - license or reuse notes - likely join key or common geography - reliability score from 1 to 5 - one reason the dataset might be inappropriate Prioritize sources from government, academic, tribal, nonprofit, or clearly documented institutional sources.
Open every promising link in the browser. Do not trust a link only because the LLM listed it.
Use the LLM as a reviewer, but make the final judgment yourself. Paste only dataset descriptions, metadata excerpts, field lists, or source summaries that you can verify in the browser.
Act as a dataset reliability reviewer for a map project. Evaluate the candidate datasets below for this map mission. Score each from 1 to 5 on: - recency - completeness - authority - granularity - bias and coverage - license clarity - field suitability - merge usefulness For each dataset, identify one claim it can support, one claim it cannot support, one likely source of bias, and one verification task a human must complete. Map mission: [PASTE] Candidate datasets: [PASTE INVENTORY OR METADATA EXCERPTS]
| Dataset | Keep / Maybe / Reject | Best Use | Main Risk | Human Verification Needed |
|---|---|---|---|---|
| Example: Short note | Example: Short note | Example: Short note | Example: Could overstate certainty | Example: Short note |
| Example: Short note | Example: Short note | Example: Short note | Example: Could overstate certainty | Example: Short note |
| Example: Short note | Example: Short note | Example: Short note | Example: Could overstate certainty | Example: Short note |
Starter examples only; expand this in your own notes or submission document.
A merged dataset should not be a pile of everything you found. Each source should have a job.
Help me assign roles to my kept datasets. Use these categories if useful: - anchor dataset: defines the rows or map features - context dataset: adds background or explanatory variables - risk or exposure dataset: adds vulnerability, hazard, or urgency - access dataset: adds services, routes, facilities, or distance context - verification dataset: checks or qualifies another source For each dataset, explain why it belongs, what field or concept it contributes, and what would happen if I removed it.
Choose one dataset to be the anchor. The anchor usually determines the final rows, such as counties, census tracts, schools, facilities, shelters, or observation records.
Create a merge plan that names the final table structure before you touch the spreadsheet.
Design a merge plan for these three kept datasets. Return: 1. final table grain: one row per [what] 2. proposed join key or matching method 3. fields to keep from each dataset 4. fields to rename for clarity 5. fields to aggregate, summarize, or classify 6. rows likely to fail matching 7. ethical or interpretive risks introduced by the merge 8. a final canonical schema with field name, definition, source dataset, and cleaning rule Datasets: [PASTE ROLES, FIELD LISTS, AND SOURCE NOTES]
Copy the canonical schema into your dossier. Do not merge until the schema is clear.
Build a small merged dataset in a browser spreadsheet. Use actual records if manageable. If a dataset is large, select a defensible subset such as one county, ten facilities, one neighborhood, or the most recent year.
Act as a data engineering assistant. Given my canonical schema and source fields, create step-by-step browser spreadsheet instructions to build the merged table. Include formulas where appropriate for XLOOKUP, VLOOKUP, FILTER, text cleanup, date parsing, category recoding, or simple scoring. Explain how to check unmatched rows and duplicate keys. Canonical schema: [PASTE] Source field notes: [PASTE] Spreadsheet tool: [Google Sheets / Excel web / other]
Prompt Cartography
canonical_dataset.Now critique the dataset you created. If you cannot upload a file to a free LLM, paste the first 12 rows, the field list, and summary notes instead.
Audit this merged task-specific dataset for map readiness. Evaluate: - missing values - duplicate keys - inconsistent categories - questionable joins - temporal mismatches - spatial mismatches - fields that imply more certainty than the sources support - sensitive or ethically risky fields - whether the dataset can support the map mission Return a problem log with severity, affected fields, likely cause, and recommended fix. Then give a publishability score from 1 to 5, but explain the score as a judgment, not a verdict.
Fix at least three issues or explain why you are choosing not to fix them.
Close the pipeline with a lineage note that a reviewer, instructor, or future collaborator could inspect.
Using my data pipeline dossier, help me draft a data lineage note with these sections: 1. map mission 2. datasets kept and why 3. datasets rejected and why 4. merge grain and join logic 5. cleaning and recoding decisions 6. reliability concerns that remain 7. claims the final dataset can support 8. claims the final dataset cannot support 9. attribution and license notes 10. next verification step before mapping Keep it concise, transparent, and suitable for inclusion with a student project or professional prototype.
Tiny nudge, large payoff: make the cartographic decision visible.
Do the tiny-but-mighty judgment check. If these answers are fuzzy, revise the map idea before polishing the prompt.
For use with Prompt Cartography: Interactive Web Map Design with LLMs, CRC Press. www.promptcartography.com