Back to skills
SKILL.md
Feature Verify
ASecuritySystem diagnosis skill (READ-ONLY). Verifies feature behavior through code analysis, data validation, and Codex deep confirmation. All operations are strictly read-only - NO data modification allowed.
- 666 stars
- 0 votes
- 0 copies
- 1 view
- Added February 8, 2026
Works with
Security analysis
100/100npx -y skills add majiayu000/claude-skill-registry --skill feature-verify --agent claude-codeAre you the author of Feature Verify?
Add the live security badge to your README. It updates with every re-scan.
[](https://www.skillsdirectory.com/skills/majiayu000-feature-verify)---
name: feature-verify
description: System diagnosis skill (READ-ONLY). Verifies feature behavior through code analysis, data validation, and Codex deep confirmation. All operations are strictly read-only - NO data modification allowed.
allowed-tools: Read, Grep, Glob, Bash, WebFetch, Task, Skill
context: fork
---
# Feature Verification Skill
## Trigger
- Keywords: verify, investigate, diagnose, check if working, anomaly, validate
- User provides credentials for data sources
## When NOT to Use
- Need to modify data (this skill is read-only only)
- Pure code review (use codex-review)
- Feature development (use feature-dev)
## Core Principle
```
ALL OPERATIONS MUST BE READ-ONLY
```
```
Claude independent analysis -> Form conclusion -> Codex third-perspective confirmation -> Integrated report
```
## Workflow
```
Phase 1: Explore -> Understand system architecture, data flow, trigger points
Phase 2: Plan -> Build verification checklist, present to user for confirmation
Phase 3: Execute -> Execute read-only queries, record expected vs actual
Phase 4: Analyze -> Claude independently forms diagnostic conclusion
Phase 5: Confirm -> /codex-brainstorm third-perspective verification
Phase 6: Integrate -> Synthesize dual perspectives, produce final report
```
## Safety Rules
| Rule | Description |
| ----------------- | ------------------------------------------------ |
| **READ-ONLY** | No write/update/delete operations allowed |
| PLAN-FIRST | Present verification plan before executing queries |
| CREDENTIAL-SAFETY | Do not expose full credentials in output |
| INDEPENDENT-FIRST | Claude forms conclusion first, then asks Codex to confirm |
## Verification Checklist
| Category | Checks |
| ------------------ | ----------------------------------- |
| Feature Checks | API endpoints, jobs, error handling |
| Data Checks | Collections, aggregations, external |
| Integration Checks | End-to-end flow |
## Verification
- Report includes Executive Summary + Status
- Each check has expected/actual/status
- Both Claude + Codex perspectives are documented
- Recommendations split into Immediate / Further / Long-term
## References
- `references/queries.md` - Query templates + safety rules
- `references/output-template.md` - Report format
## Examples
```
Input: /feature-verify User Auth
MongoDB: mongodb://...
{ANALYTICS}: project_id={PROJECT_ID}
Action: Explore -> Plan -> Execute queries -> Analyze -> Codex confirm -> Report
```
```
Input: verify user auth is working correctly
Action: Phase 1-6 flow -> Output diagnostic report
```
Attribution
Comments
Loading comments…
Newsletter
Get the weekly list of safe new skills, and the ones we pulled.