Oracle Confidentiality

Oracle Confidentiality Guardian

> Purpose: Enforce confidentiality protocol across all Oracle AI Architect deliverables. > Scope: Pre-delivery audit, codename enforcement, content sanitization. > Authority: This skill has VETO power. If it fails, delivery is blocked.

When to Use This Skill

Invoke:

• Before delivering ANY document to a customer
• Before committing client-folder content to git
• When generating content that references customer context
• As the final step of /oracle-solution-design Phase 5

Trigger: /oracle-confidentiality or automatically at end of solution design workflow

---

The Codename Protocol

Rules (Non-Negotiable)

1. Codenames are OPAQUE -- A, B, E, K, O, P, R, V have no inherent meaning
2. Never persist context -- Industry, scope, employee count, revenue NEVER in committed files
3. Conversation-only context -- Tell Claude what the codename means at session start, it stays in memory only
4. README.md is the only committable file per client -- contains ONLY: status, role, codename
5. clients/.gitignore blocks: deliverables/, notes/, docs/, SOLUTION-DESIGN.md, auto-CLAUDE.md
6. Research goes to research/topics/ -- NEVER to research/projects/[CODE]/ (would link codename to topic)

What CAN Be in Committed Files

• Codename letter only (A, B, K, etc.)
• Status (Active, Prospect, Completed)
• Role (AI/Cloud Architecture)
• Generic dates and milestones

What MUST NEVER Be in Committed Files

• Real customer name
• Industry vertical
• Geographic region or country
• Employee count or revenue
• Contract value or pricing
• Specific technology stack of the customer
• Names of customer employees
• Internal Oracle pricing or discounts

---

Pre-Delivery Audit Checklist

Run this checklist before any deliverable leaves the workspace:

Step 1: Content Scan

Search all output files for:
- Real customer names (from conversation context)
- Industry-specific terms that could identify the client
- Geographic identifiers tied to the client
- Internal Oracle pricing not on public price list
- Competitor names used in attack mode (not just comparison)

Step 2: File Location Check

Verify:
- Deliverables are in clients/[CODE]/deliverables/ (gitignored)
- SOLUTION-DESIGN.md is in clients/[CODE]/ (gitignored)
- No deliverables leaked to research/ or projects/ folders
- No codename appears in research/topics/ filenames

Step 3: Git Safety Check

Before any commit:
- Run: git diff --cached -- check no client content staged
- Run: git status -- verify clients/ content is untracked
- Verify clients/.gitignore is intact and blocking deliverables

Step 4: Image Compliance

For every generated image:
- No Oracle logos (text labels only)
- No customer logos
- No identifiable customer branding or colors
- Service names match official Oracle branding

Step 5: Document Sanitization

In every document:
- Customer referred to as "the organization" or "the customer"
- Solution name is generic or codename-based
- All data examples use synthetic/mock data
- No internal meeting notes or email quotes

---

Automated Checks (for Agent Execution)

When invoked, the agent should execute these checks:

1. Grep for known risks:

• Grep output files for any real names mentioned in conversation
• Grep for currency amounts (could indicate contract values)
• Grep for specific addresses or locations

1. Verify .gitignore integrity:

• Read clients/.gitignore
• Confirm it blocks: deliverables/, notes/, docs/, SOLUTION-DESIGN.md

1. Check git status:

• Ensure no client deliverables are staged or tracked
• Warn if any new files in clients/ are untracked but not gitignored

1. Report:

• PASS: All checks passed, safe to deliver
• FAIL: List specific violations with line numbers and file paths
• WARN: Potential issues that need human review

---

Emergency Protocol

If confidential data is accidentally committed:

1. DO NOT push
2. Soft reset: git reset HEAD~1 (undo last commit, keep files)
3. Remove sensitive content from files
4. Re-commit with clean content
5. If already pushed: Contact user immediately, may need force push (with user approval)

---

Integration

This skill is called by:

• /oracle-solution-design (Phase 5: DELIVER)
• Any manual invocation before delivery
• Should be invoked proactively by Claude when generating client-facing content

---

Version: 1.0 | Created: 2026-02-09

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: oci-ai-architects
• Source: oci-ai-architects/claude-code-oci-ai-architect-skills
• License: MIT
• Homepage: https://oci-ai-architects.github.io/claude-code-oci-ai-architect-skills

Install and usage instructions live in the source repository linked above.