lee-fuhr Part of Claude Code operator skills: a collection of skills for running a real Claude Code setup.
Audit data quality with expert personas
Audit data quality by running 15 expert personas in serial, fixing critical issues per framework before verifying.
15 data quality frameworks, each loaded as a deep expert persona into its own agent context. Serial execution: one framework → fix critical issues → verify → next. Each agent thinks like a 20-year specialist in that specific framework.
Expert lens: Is the data trustworthy? Are decisions being informed?
Supersedes: ad-hoc analytics reviews, one-off tracking audits.
Modes
This skill responds to three modes based on the args passed. Interpret intent loosely, using fuzzy matching, not exact phrasing.
Mode 1: Full serial audit (no args or "run all")
Triggers: /qq-audit-data, "run all", "full audit", "everything", no args at all.
Mode 2: Single framework (framework name + optional scope)
Triggers: /qq-audit-data funnel instrumentation, "just run dashboard accuracy on the metrics page", "check event taxonomy", etc.
Match the framework name fuzzily: "funnel", "conversion", "drop-off" should all match Funnel Instrumentation. "dashboard", "metrics", "accuracy" should all match Dashboard Accuracy Audit. If ambiguous, show the 2-3 closest matches and ask.
Mode 3: List frameworks (help/list/discovery)
Triggers: "list", "what do you have", "what frameworks", "help", "which ones", "show me", etc.
Show the framework table from the framework inventory below, then ask which one(s) to run.
Smart interview (runs before any audit)
DO NOT ask dumb questions. Before asking anything, gather what you already know:
- Check conversation context: what product are we working on? What files have been discussed? What has the user been complaining about?
- Check project CLAUDE.md: product description, tech stack, target audience.
- Check recent session state: what was just built or changed?
Pre-fill and present assumptions:
"Here’s what I know going in:
- Product: [name], [description from context]
- Analytics stack: [GA4 / Mixpanel / PostHog / custom, inferred from project]
- Data layer: [GTM / Segment / custom dataLayer, inferred from code]
- Known data concerns: [what the user has mentioned or what recent changes suggest]
- Scope: [full tracking / specific funnel / specific dashboard]
Anything wrong or missing?"
Use AskUserQuestion with multiple choice ONLY for genuine gaps. For example, if you truly can’t tell what analytics platform is in use, ask. If you can infer it, state the inference.
Interview output becomes the audit context: passed to every framework agent so they audit with purpose, not generically.
Execution flow (full audit)
Phase 1: Context gathering
- Run smart interview (above)
- Read the product’s analytics config, event tracking code, and data layer implementation to understand scope
- Map the data surface: events tracked, funnels defined, dashboards deployed, third-party integrations
Phase 2: Serial framework execution
For each framework (in order 1-15):
Spawn one agent with these instructions:
- "Read the framework subskill file at
~/.agents/skills/qq-audit-data/frameworks/[NN-name].md" - "You ARE the expert described in that document. Think like them, not like a generalist."
- "Audit [product name] focusing on [scope]. Context: [interview output]"
- "Previous frameworks found these issues: [cumulative findings list]. Do NOT re-report duplicates."
- "Score 1-10, list findings with file:line references, fix critical issues (score < 7), write report."
- "Read the framework subskill file at
Collect the agent’s output:
- Score for this framework
- Findings (new, not duplicates)
- Fixes applied
- Items flagged for downstream frameworks
Verify build if fixes were applied (
npx next buildor equivalent)Update cumulative state:
- Add findings to the dedup list
- Update the running score card
- Note cross-framework flags for upcoming frameworks
Move to next framework
Phase 3: Synthesis
- Run framework #15 (Real-Time vs Batch Freshness) as the final timeliness review
- Compile consolidated report:
- Overall data quality score
- Per-framework scores
- All findings grouped by severity (critical / high / medium / low)
- Fixes applied (with file:line)
- Remaining items for future rounds
- Compare to previous audit rounds if they exist
Phase 4: Report
Save to project’s data directory:
data/audit-data-[date].md: full reportdata/audit-data-[date]-summary.md: scores + critical findings only
Framework inventory
| # | Framework | Expert lens | Subskill file |
|---|---|---|---|
| 1 | Analytics Implementation Completeness | Measurement coverage: are you tracking every key user action, funnel step, and business event that drives decisions? | 01-analytics-completeness.md |
| 2 | Data Layer Architecture | Data plumbing: is event data flowing through a clean, centralized data layer instead of scattered vendor-specific calls? | 02-data-layer.md |
| 3 | Event Taxonomy Consistency | Naming discipline: do events follow a consistent, documented naming convention that makes querying intuitive and reliable? | 03-event-taxonomy.md |
| 4 | Data Validation and Integrity | Data correctness: do integrity constraints, validation rules, and automated checks catch data corruption before it reaches dashboards and decisions? | 04-data-validation.md |
| 5 | Privacy-Compliant Tracking | Consent and privacy: does tracking respect user choices, anonymize where required, and comply with GDPR/CCPA requirements? | 05-privacy-compliant-tracking.md |
| 6 | Funnel Instrumentation | Conversion path visibility: does every critical funnel have events at every step with drop-off visibility between steps? | 06-funnel-instrumentation.md |
| 7 | A/B Testing Infrastructure | Experimentation rigor: can you run controlled experiments with proper randomization, sufficient sample sizes, and statistically valid conclusions? | 07-ab-testing-infra.md |
| 8 | Data Retention and Lifecycle | Data lifecycle hygiene: are retention periods defined, automated cleanup running, and old data properly archived or deleted? | 08-data-retention.md |
| 9 | Dashboard Accuracy Audit | Metric trustworthiness: do dashboard numbers match their source data, are calculations correct, and can stakeholders trust what they see? | 09-dashboard-accuracy.md |
| 10 | Error Tracking Coverage | Failure visibility: are client and server errors captured with enough context to diagnose the problem and understand its user impact? | 10-error-tracking.md |
| 11 | Search Analytics | Search intelligence: are search queries tracked, zero-results visible, and search quality measured so you know what users want but can’t find? | 11-search-analytics.md |
| 12 | Attribution Modeling | Conversion credit: does your attribution model match business reality, or is it giving credit to the wrong channels? | 12-attribution-modeling.md |
| 13 | Data Export and Portability | Data freedom: can users get their data out in standard formats, and is the system a partner rather than a data roach motel? | 13-data-export.md |
| 14 | Schema Evolution Management | Schema change safety: can you add, modify, or remove event properties and data fields without breaking existing queries, dashboards, and pipelines? | 14-schema-evolution.md |
| 15 | Real-Time vs Batch Freshness | Data timeliness: does data freshness match business needs, and is the freshness clearly labeled so stakeholders know what they’re looking at? | 15-data-freshness.md |
Key principles
- Serial, not parallel: 70% of findings duplicate across frameworks. Serial means each round finds genuinely new issues after fixes.
- Fix before moving on: don’t accumulate a findings list. Fix each framework’s criticals before the next audit.
- Expert persona, not checklist: each agent IS the specialist. They reason from principles, not rules.
- Hold every fix to a real quality bar: is this real data? Does it prevent errors? Is the complexity earned?
- Code + live data: code audits miss data quality issues. Always check actual event payloads, dashboard outputs, and query results when possible.
- Multi-round: after all 15 frameworks, run the full cycle again. Scores increase each round until plateau.
- Dedup across frameworks: each agent receives cumulative findings so they don’t re-report known issues.