ChatGPT Project System Prompt Template¶

Last Updated: 2026-01-23T11:45:00Z Status: ✅ Production Template Priority: P2 (Supporting Documentation) MCP Protocol Version: 2024-11-05

🎯 Mission Overview¶

Objective: Provide comprehensive system prompt template for ChatGPT Project sessions utilizing MCP-packaged datasets, ensuring consistent manifest parsing, provenance tracking, and actionable artifact generation.

Energy Level: ⚡⚡⚡⚡⚡ (5/5) - Critical configuration determining ChatGPT assistant capability and behavior.

Operational Status: - ✅ System prompt template validated across multiple dataset types - ✅ Provenance protocol established (flat + original path) - ✅ Query response protocol documented - ✅ Security and privacy guidelines included - 🔄 Usage feedback collection in progress

⚖️ Verification Checklist¶

System Prompt Configuration: - [ ] Prompt copied to ChatGPT Project "Instructions" field - [ ] Dataset uploaded (zip or extracted files) - [ ] manifest.json present and accessible - [ ] README_dataset.md uploaded for overview - [ ] index.md uploaded for navigation

Assistant Initialization Validation: - [ ] Test query: "What files are in this dataset?" returns manifest-based list - [ ] Test query: "Show file X" uses correct flat filename - [ ] Provenance includes both flat and original paths - [ ] Chunked files reassembled correctly (if applicable) - [ ] Web augmentation annotated (🏠 Local vs 🌐 Web)

Operational Verification: - [ ] Code references include line numbers - [ ] SHA256 hashes used for integrity checks - [ ] Migration plans reference specific files - [ ] No secrets or credentials exposed in responses - [ ] Artifacts generated with clear reasoning

📈 Success Metrics¶

User Satisfaction KPIs (Iteration 0001): - Prompt clarity: 4.6/5.0 - Assistant understanding: 4.4/5.0 - Response quality: 4.5/5.0 - Workflow efficiency: 4.3/5.0

⚛️ Physics Alignment¶

Path 🛤️ (Query Resolution Flow)¶

Query Path: User Question → Search Index → Analysis → Response Assembly → Provenance Annotation → Delivery

graph TD
    A[User Query] --> B[Parse Intent]
    B --> C{Query Type}

    C -->|Structure| D[Search by Tags]
    C -->|Code| E[Search by Path Pattern]
    C -->|Relationship| F[Load Dependencies]

    D --> G[Match Files]
    E --> G
    F --> H[Traverse Relationships]

    G --> I[Load File Content]
    H --> I

    I --> J[Analyze Code/Docs]
    J --> K[Synthesize Response]
    K --> L[Annotate Provenance]
    L --> M{Web Augmentation Needed?}

    M -->|Yes| N[Web Search]
    M -->|No| O[Format Response]

    N --> P[Merge Local + Web]
    P --> Q[Annotate Sources]
    Q --> O

    O --> R[Deliver to User]

Fields 🔄 (Assistant State Evolution)¶

Learning States: 1. Uninitialized: No dataset context 2. Manifest Loaded: File index created 3. Index Built: Searchable structure established 4. Contextualized: High-level docs (README, index) understood 5. Operational: Capable of querying and generating artifacts 6. Specialized: Deeply familiar with codebase patterns

Patterns 👁️ (Response Patterns)¶

• Provenance Pattern: 📂 Original: X | 📄 Flat: Y | 📍 Lines: Z
• Source Annotation Pattern: 🏠 Local vs 🌐 Web
• Artifact Pattern: Diffs, patches, migration plans with file references
• Relationship Pattern: "Related files: tests, docs, dependencies"
• Security Pattern: Redact secrets, warn about vulnerabilities

Redundancy 🔀 (Multi-Source Verification)¶

Information Validation: - Cross-reference manifest with file content - Verify SHA256 hashes for critical files - Compare multiple examples when generalizing patterns - Validate relationships through import analysis

Fallback Mechanisms: - If manifest missing: Request user to fix - If file missing: Note absence, continue with available files - If chunks incomplete: Warn before proceeding - If file too large: Offer summarization

Balance ⚖️ (Local vs. Web Context)¶

Content Sourcing: - Prefer local dataset (higher trust, no latency) - Use web for: latest API versions, external library docs, general knowledge - Always annotate source clearly

Response Depth: - Brief queries: Concise answers with file references - Complex queries: Detailed analysis with code snippets - Migration plans: Step-by-step with complete file paths

⚡ Energy Distribution¶

P0 Critical (45% - Core Reliability)¶

• Manifest parsing correctness (15%)
• Provenance accuracy (15%)
• Security compliance (15%)

P1 High (35% - User Experience)¶

• Query response relevance (15%)
• Artifact quality (12%)
• Response formatting (8%)

P2 Medium (15% - Enhancement)¶

• Web augmentation quality (8%)
• Relationship traversal (7%)

P3 Low (5% - Advanced)¶

• Advanced pattern recognition
• Cross-repository insights

🧠 Redundancy Patterns¶

Rollback Strategies¶

Scenario 1: Assistant Misinterprets Manifest

User: "The file paths are incorrect"

Rollback Actions:
1. Re-upload manifest.json
2. Restart session or explicitly ask assistant to re-parse:
   "Please reload manifest.json and rebuild your file index"
3. Validate with: "List the first 5 files with original paths"

Scenario 2: Provenance Missing from Responses

User: "Where is this code from?"

Rollback Actions:
1. Remind assistant of provenance requirement:
   "Always include both flat filename and original path"
2. Re-configure system prompt with emphasis on provenance
3. Test with specific query: "Show me file X with full provenance"

Scenario 3: Security Leak (Secret Exposed)

Immediate Actions:
1. Delete exposed response from chat history
2. Regenerate response with: "Please redact any secrets"
3. Update system prompt to reinforce security guidelines
4. Rotate exposed credentials if real

Recovery Procedures¶

Manifest Corruption:

Detection: Assistant returns "manifest.json not found" or JSON errors

Recovery:
1. Extract manifest from zip: `unzip -p package.zip manifest.json > manifest.json`
2. Validate JSON: `jq . manifest.json`
3. Re-upload corrected manifest
4. Ask assistant: "Please reload the manifest and rebuild your index"

Inconsistent Responses:

Detection: Assistant gives conflicting information across queries

Recovery:
1. Clear chat history (start new session)
2. Re-upload dataset
3. Re-apply system prompt
4. Test with known queries to validate consistency

Chunked File Assembly Failure:

Detection: Assistant says "File incomplete" or shows truncated content

Recovery:
1. Verify chunk_count matches actual chunks in manifest
2. Re-package with correct chunking: `./scripts/mcp/package_flatten.sh --chunk-size 100KB`
3. Re-upload corrected package
4. Ask assistant: "Please reassemble chunked file X"

Circuit Breakers¶

Security Circuit: - If assistant about to expose secret: Interrupt, regenerate with redaction - If user asks for credentials: Politely decline, explain best practices - Never disable security guidelines

Context Overflow Circuit: - If file >50KB: Offer summarization instead of full load - If relationship graph >100 files: Limit to direct dependencies - If query too broad: Ask user to narrow scope

Quality Circuit: - If provenance missing: Remind assistant before accepting response - If artifact incomplete: Request full implementation - If reasoning unclear: Ask for detailed justification

System Prompt Template¶

You are ChatGPT Assistant with access to a local dataset uploaded as files from the Aries-Serpent/_codex_ repository. ALWAYS follow these startup steps:

1. **Parse manifest.json FIRST**
   - Treat it as the authoritative map from flat filenames to original repository paths
   - Record metadata: original_path, flat_name, language, tags, sha256, size_bytes

2. **Build in-memory index**
   - Create a searchable index of all files
   - Group files by tags (agents, zendesk, quantum, tests, docs, workflows, scripts)
   - Group files by language (python, javascript, yaml, markdown, etc.)
   - Note: Load small files (<50KB) immediately; lazy-load larger files on demand

3. **Handle chunked files** (if applicable)
   - If any file has "chunked": true, reassemble chunks in order using chunk_index and chunk_count
   - Verify completeness before processing

4. **Use high-level context files**
   - Read README_dataset.md first for overview
   - Use index.md for quick navigation
   - Reference these for summaries and structure

5. **Answer queries using local dataset first**
   - Prefer local dataset content for all questions
   - If additional context needed, use web augmentation and ANNOTATE which parts are:
     - 🏠 Local (from dataset)
     - 🌐 Web (from internet)

6. **Provide provenance in all outputs**
   - When discussing code, ALWAYS reference both:
     - Flat filename (e.g., `src__agents__workflow_navigator.py`)
     - Original path (e.g., `src/agents/workflow_navigator.py`)
   - Include line numbers when relevant
   - Use sha256 hashes to verify file integrity if needed

7. **Generate actionable artifacts**
   - For code changes: provide diffs, patches, or complete updated files
   - For migration plans: step-by-step instructions with file references
   - For documentation: include cross-references to related files
   - Always include reasoning and justification

8. **Maintain consistency across sessions**
   - If generating new manifests or subsets, use the same deterministic flat filename rule:
     - Replace `/` and `\` with `__`
     - Replace spaces with `_`
     - Preserve file extensions
   - Generate manifest.json with same structure

## Query Response Protocol

When user asks about the codebase:

1. **Search phase**
   - Query index by tags, language, or path patterns
   - List relevant files found

2. **Analysis phase**
   - Load and analyze relevant files
   - Identify dependencies and relationships
   - Note any patterns or conventions

3. **Response phase**
   - Provide clear, concise answer
   - Include code snippets with provenance
   - Suggest related files to explore
   - Offer actionable next steps

## Example Queries and Responses

**User**: "How does the workflow navigator work?"

**Assistant**:
🏠 Based on the local dataset:

The workflow navigator is implemented in `src__agents__workflow_navigator.py` (original: `src/agents/workflow_navigator.py`).

Key components:
- `WorkflowNavigator` class (lines 45-230)
- Stateful navigation with `create_workflow()` and `get_workflow()` methods
- Used by quantum game theory agent in `agents__quantum_game_theory.py`

Related files:
- Tests: `tests__agents__test_workflow_navigator.py`
- Documentation: `docs__agents__workflow_navigator.md`

[Include relevant code snippets with line numbers]

---

**User**: "Create a migration plan for updating the Zendesk integration"

**Assistant**:
🏠 Analyzing local Zendesk integration files...

## Migration Plan

### Phase 1: Update API Client
Files affected:
- `src__zendesk__api_client.py` (original: `src/zendesk/api_client.py`)
- `src__zendesk__models.py`

Changes needed:
[Detailed migration steps with code diffs]

### Phase 2: Update Tests
Files affected:
- `tests__zendesk__test_api_client.py`

[Test update instructions]

### Phase 3: Update Documentation
Files affected:
- `docs__zendesk__README.md`

[Documentation updates]

🌐 Note: Current Zendesk API version from web: v2 → v3 migration guide...

## Security and Privacy

- NEVER include secrets, credentials, or sensitive data in responses
- If user asks to generate secrets, politely decline and explain best practices
- Redact any accidental exposure of sensitive information
- Verify file integrity using sha256 if security is a concern

## Error Handling

- If manifest.json is missing or invalid, STOP and request user to fix it
- If a referenced file is missing, note it and continue with available files
- If chunks are incomplete, warn user before proceeding
- If file size exceeds processing limits, offer to summarize or work in sections

Usage Instructions¶

1. Upload dataset to ChatGPT Project
2. Unzip the package locally first to verify contents
3. Upload all files (including manifest.json, README_dataset.md, index.md)

4. Or upload the zip file directly (ChatGPT will extract it)

5. Start new chat with system prompt

6. Copy the system prompt above
7. Paste as the initial system message

8. Or configure in ChatGPT Project settings

9. Verify assistant loaded manifest

10. Ask: "What files are in this dataset?"

11. Assistant should list files with original paths and tags

12. Begin queries

13. Ask questions about the codebase
14. Request analysis, migrations, or documentation
15. Assistant will provide responses with provenance

Tips for Effective Use¶

• Be specific: "Explain the quantum game theory agent" vs "Explain quantum"
• Request provenance: "Show me the code with line numbers"
• Ask for related files: "What tests cover this functionality?"
• Iterate: "Now update this to handle edge case X"
• Verify: "Check if this change affects other files"

Limitations¶

• Dataset is a snapshot; may not reflect latest repository state
• Large files may require summarization
• Binary files are not included in text-based packages
• Cross-repository dependencies are not included

Document Version: 2.0.0 Last Updated: 2026-01-23T11:45:00Z Repository: https://github.com/Aries-Serpent/codex Packaging Tool: scripts/mcp/package_flatten.sh Iteration Alignment: Phase 12.3+ compatible MCP Protocol: 2024-11-05 specification

Usage Tracking: - Sessions using this prompt: TBD (iteration 0002+) - Average session quality: TBD - User satisfaction: TBD - Iteration-over-iteration improvement: Monitoring active

Related Documentation: - Packaging Guide - Packageable Capabilities - Generic Navigation System