productize-explore-data
Profile and explore a dataset to understand its shape, quality, and patterns. Use when encountering a new table or file, checking null rates and distributions, spotting duplicates or suspicious values, or deciding which dimensions and metrics to analyze.
- License 许可
- MIT
- 最近更新
- 2026-08-07
- Stars
- ★ 3
- Forks
- 0
01 前言 Front Matter
- name
- productize-explore-data
- description
- Profile and explore a dataset to understand its shape, quality, and patterns. Use when encountering a new table or file, checking null rates and distributions, spotting duplicates or suspicious values, or deciding which dimensions and metrics to analyze.
02 资源文件 1
- 参考文档agents/openai.yaml325 B
03 SKILL.md 正文
<!-- AUTO-GENERATED from SKILL.md.tmpl - do not edit directly --> <!-- Regenerate: npm run skills:generate -->
Explore Data
Productize Preamble
Before producing the artifact, implementation step, or code change, classify the work:
- User/persona: founder, product leader, AI PM, AI builder, stakeholder, or mixed/unknown.
- Product stage: idea, validation, PMF search, growth, scale, pivot, or unknown.
- Artifact mode: strategy memo, PRD, research plan, positioning, experiment, deck, roadmap, execution brief, diagnostic, or decision record.
- Artifact format: Markdown for short, repo-native, diff-sensitive artifacts; self-contained HTML for long, visual, shareable, interactive, or explicitly requested artifacts.
- Evidence standard: what is known, assumed, missing, and risky.
- Decision mode: recommend, ask for a blocking input, or proceed with explicit assumptions.
Operating rules:
- Do not produce generic product strategy filler. Tie every recommendation to the user's context, evidence, stage, and decision pressure.
- Separate facts from assumptions. Convert high-risk assumptions into tests, research prompts, or instrumentation.
- Search existing context, attached docs, repo files, or prior Productize artifacts before inventing new structure when those sources are available.
- Ask only for inputs that materially change the output. If the next step is obvious, proceed and state assumptions.
- If another Productize skill is a better fit, name it and explain the routing in one sentence before continuing.
- Enforce the output contract for this skill. Produce the artifact, implementation step, verification evidence, or code change; do not stop at a meta-explanation of the method.
- End with the concrete next action, decision owner, validation step, metric, or artifact handoff when applicable.
- Use the user's saved Productize preferences when available, but do not let stale preferences override explicit instructions in the current request.
- If you must ask a blocking question, use this compact format:
AskUserQuestion: <one question> Why it matters: <decision it changes>. - Record completion status for durable work as
completed,blocked,deferred, orneeds-reviewwhen runtime hooks are available.
Artifact format policy:
- Default to Markdown for short notes, repo-native documentation, changelog fragments, or artifacts where clean source diffs matter.
- Use a self-contained HTML file when the user asks for HTML, the artifact is likely to exceed about 100 lines, the reader needs diagrams/tables/comparisons/screenshots, the artifact is meant to be shared, or interaction/export controls would help the user stay in the loop.
- Route by product job first, then choose the format. Do not create a generic HTML workflow when a Productize playbook, gate, or routed skill owns the work.
- HTML artifacts must be local-first and portable: one file, embedded CSS/JS, no remote dependencies unless explicitly requested, readable without a dev server, responsive, accessible, and easy to skim.
- Prefer semantic sections, tables, SVG diagrams, annotated code blocks, status chips, collapsible detail, and copy/export buttons when they improve review speed. Avoid decorative complexity that hides the decision.
- For implementation work with ambiguities, use
$implementation-noteswhen the user asks for a running notes file, especiallyimplementation-notes.htmlorimplementation-notes.md.
Runtime hooks, if available:
- Use
productize-workflow start "<user request>"at the beginning of durable product work; it restores context, routes the request, records the session, and returns the required artifact contract. - Use
productize-workflow complete --id <id> --status completed|blocked|deferred|needs-review --artifact-type <type> --summary <summary>before ending durable product work; it records the artifact, saves context, and logs completion status. - Use
productize-update-check --strictat the start of maintenance, setup, release, or generated-output work. - Use
productize-configto read or write user/team preferences such as persona, artifact mode, evidence threshold, or update-check policy. - Use
productize-session-logto record important workflow decisions. - Use
productize-artifact-logwhen a durable artifact is produced. - Use
productize-context-restorebefore restarting long-running product work from scratch. - Use
productize-context-saveafter producing a durable strategy, research, spec, or stakeholder artifact. - Use
productize-registry-searchorproductize-skill-routerwhen routing is ambiguous. - Use
productize-completion-statusto log whether the workflow completed, blocked, deferred, or needs review.
Telemetry standard:
- Keep telemetry local by default in
.productize/orPRODUCTIZE_STATE_DIR. - Log artifact type, routing decision, evidence gaps, and completion status; do not log secrets or private customer data unless the user explicitly asks for it.
Current skill metadata:
- Skill:
explore-data - Lifecycle: Measure
- Category: Analytics
- Primary artifact: Explore Data analytics diagnosis with metric readout, caveats, decision, and next instrumented step
Generate a comprehensive data profile for a table or file before deeper analysis.
Argument Hint
<table or file>
Usage
/explore-data <table_name or file>
Workflow
1. Access Data
If a data warehouse or analytics MCP is connected:
- Resolve the table name, including schema prefixes.
- Query table metadata: columns, types, descriptions, size, update metadata if available.
- Run profiling queries against live data. Use sampling for very large tables and disclose it.
If a file is provided:
- Load CSV, Excel, Parquet, or JSON into a working dataset.
- Infer column types and normalize obvious parsing issues.
If neither is available, ask for a table name, file, pasted sample, or schema description. If only a schema is provided, give profiling SQL the user can run.
2. Understand Structure
Answer table-level questions:
- Rows and columns.
- Grain: one row per what.
- Primary key or natural key, and whether it is unique.
- Last updated timestamp and expected refresh cadence if known.
- Date coverage and earliest/latest records.
Classify columns:
- Identifier: primary keys, foreign keys, entity IDs.
- Dimension: categorical attributes for grouping/filtering.
- Metric: quantitative values.
- Temporal: dates and timestamps.
- Text: free-form strings.
- Boolean: true/false flags.
- Structural: JSON, arrays, nested data.
3. Profile Columns
Table-level:
- Total row count.
- Column count and type breakdown.
- Approximate table size if available.
- Date range for temporal columns.
All columns:
- Null count and null rate.
- Distinct count and cardinality ratio.
- Top 5-10 most common values with frequencies.
- Rare values when useful for anomaly detection.
Numeric columns:
- Min, max, mean, median, standard deviation.
- Percentiles: p1, p5, p25, p75, p95, p99.
- Zero count and negative count.
- Distribution shape and outliers.
String columns:
- Min, max, and average length.
- Empty string count.
- Pattern consistency.
- Case consistency.
- Leading/trailing whitespace.
Temporal columns:
- Min and max dates.
- Null and future dates.
- Distribution by month/week.
- Time series gaps.
Boolean columns:
- True count, false count, null count, true rate.
4. Flag Data Quality Issues
Use severity labels: high, medium, low.
Flag:
- High null rates: warn above 5 percent, alert above 20 percent.
- Duplicate natural keys.
- Low-cardinality IDs or unexpectedly high-cardinality categorical fields.
- Suspicious values: negative amounts, future dates, placeholders, test values, impossible ranges.
- Skewed distributions that make averages misleading.
- Encoding and formatting issues: mixed case, whitespace, inconsistent formats.
- Cross-column rule violations: completed status with missing completed_at, end date before start date.
- Staleness or unexpected load lag.
5. Discover Patterns
Look for:
- Foreign key candidates.
- Natural hierarchies such as country -> state -> city.
- Numeric correlations worth investigating.
- Derived columns that appear calculated from other fields.
- Redundant or near-identical columns.
- Natural segments based on categorical fields with useful cardinality.
6. Recommend What To Analyze Next
Suggest:
- Best dimensions for slicing data.
- Key metric columns.
- Time columns suitable for trends.
- Natural groupings or hierarchies.
- Potential join keys.
- 3-5 concrete follow-up analyses.
Examples:
- Trend analysis on
[metric]by[time_column]grouped by[dimension]. - Distribution deep dive on
[skewed_column]. - Data quality investigation on
[problematic_column]. - Correlation analysis between
[metric_a]and[metric_b]. - Cohort analysis using
[date_column]and[status_column].
Output Format
## Data Profile: [table_or_file]
### Overview
- Rows:
- Columns:
- Grain:
- Primary key:
- Date range:
- Last updated:
### Column Details
[Summary table grouped by IDs, dimensions, metrics, dates, booleans, text, structural fields]
### Data Quality Issues
[Severity, field, issue, evidence, recommendation]
### Patterns and Relationships
[Join keys, hierarchies, correlations, derived/redundant fields]
### Recommended Explorations
[Numbered list of follow-up analyses]
Quality Framework
Completeness:
- Complete: more than 99 percent non-null.
- Mostly complete: 95-99 percent.
- Incomplete: 80-95 percent.
- Sparse: below 80 percent.
Consistency checks:
- Same concept represented multiple ways.
- Numbers stored as strings.
- Dates in inconsistent formats.
- Foreign keys without parents.
- Business rule violations.
Accuracy red flags:
- Placeholder values such as 0, -1, 999999, N/A, TBD, test, xxx.
- Suspicious default values.
- Stale active-system data.
- Impossible values.
- Round-number bias.
Timeliness:
- Last updated time.
- Expected update frequency.
- Event-to-load lag.
- Gaps in expected time series.
Schema Documentation Template
When documenting the dataset for team use, include:
- Description.
- Grain.
- Primary key.
- Row count and date.
- Update frequency.
- Owner if known.
- Key columns table.
- Relationships.
- Known issues.
- Common query patterns.
SQL Discovery Patterns
Use dialect-appropriate schema queries. For PostgreSQL-style systems:
SELECT table_name, table_type
FROM information_schema.tables
WHERE table_schema = 'public'
ORDER BY table_name;
SELECT column_name, data_type, is_nullable, column_default
FROM information_schema.columns
WHERE table_name = 'my_table'
ORDER BY ordinal_position;
For other warehouses, adapt to the warehouse dialect and metadata tables.
04 同仓库更多 Skills
workflow-memory
Maintains workflow-scoped task memory for Productize runs using .productize/tasks/<name>/memory/ files. Use when a task prompt provides workflow memory paths and requires the agent to read, update, compact, and promote durable context across PRD task executions. Do not use for PR review remediation, global user preferences, or programmatic event-log summarization.
review-round
Performs a comprehensive code review of a PRD implementation and generates a review round directory with issue files compatible with fix-reviews. Use when reviewing implemented PRD tasks, creating a manual review round without an external provider, or performing a quality audit of code changes. Do not use for fetching reviews from external providers, fixing existing review issues, executing PRD tasks, or editing source code.
productize
Productize root skill for building, shipping, operating, and growing AI products inside a coding harness. Routes work across thesis framing, discovery, research, positioning, finance, decision making, requirements, design, metrics, AI-builder execution, code implementation, verification, stakeholder narratives, growth, launch, operation, and learning workflows. Use when the user wants to build a product from zero, improve an existing product, or needs product work that spans multiple lifecycle stages before code, launch, operation, or growth execution.
productize-wwas
Why-What-Acceptance. Use when writing backlog items in Why-What-Acceptance format, breaking features into independent valuable work, or adding strategic context to delivery items.
productize-writing-plans
Create comprehensive implementation plans with bite-sized tasks. Use when you have requirements before touching code.
productize-write-spec
Write a feature spec or PRD from a problem statement or feature idea. Use when turning a vague idea into a structured document, scoping goals and non-goals, defining success metrics and acceptance criteria, or breaking a big ask into phases.