Skip to main content

Model-Specific Guidance

Manicode prompts are available in variants optimized for specific models. This page documents practical differences across supported models and guidance for choosing between them.

Supported Models

VendorModelPrompt Variant SuffixPrimary Strengths
AnthropicClaude Opus 4.6(Opus 4.6)Long-form reasoning, instruction adherence, security analysis
OpenAIGPT 5.3 Codex(GPT 5.3 Codex)Code generation, structured output, tool use
GoogleGemini 3.1 Pro(Gemini 3.1 Pro)Long context windows, multimodal input, engineering tasks
xAIGrok 4.1(Grok 4.1)Fast inference, code generation
MicrosoftGitHub Copilot(Copilot)IDE-integrated code completion, inline suggestions

When to Use Which Model

Code Security Prompts

All five model variants are available for code security prompts. Choose based on your deployment context:

Deployment ContextRecommended ModelReason
IDE autocompleteGitHub CopilotNative IDE integration, inline completions
Coding agent (autonomous)Claude Opus 4.6, GPT 5.3 CodexStrong instruction following for long autonomous sessions
API-driven code generationAny modelAll variants produce comparable security-aware code
Code review / analysisClaude Opus 4.6, Gemini 3.1 ProDetailed reasoning in review comments

Threat Modeling Prompts

Threat modeling prompts are available for Claude Opus 4.6, GPT 5.3 Codex, and Gemini 3.1 Pro.

MethodologyRecommended ModelReason
Standard InterviewClaude Opus 4.6Patient, methodical multi-turn interviews
STRIDE for APIsGPT 5.3 CodexStructured, systematic STRIDE analysis
FMEAGemini 3.1 ProEngineering-focused quantitative analysis
Diagram GenerationAny supported modelMermaid/PlantUML output is consistent across models
Repository ReconnaissanceClaude Opus 4.6Best file navigation and cross-referencing in agent mode

Compliance Prompts

Compliance prompts do not have model-specific variants. Use any supported model. For best results:

  • Long intake sessions — Use a model with a large context window (Gemini 3.1 Pro, Claude Opus 4.6)
  • Structured output — If you need JSON-formatted control mappings, GPT 5.3 Codex handles structured output well

Validation Prompts

Validation prompts are designed to be model-agnostic. Choose based on latency and cost requirements:

RequirementRecommended Approach
Low latency (production guardrails)Use Low Token variants with a fast model
High accuracy (audit/development)Use High Token variants with a larger model
High throughputUse the smallest model that meets your accuracy threshold

Workforce and AI Coding Requirements

These prompts are available for Claude Opus 4.6 and GPT 5.3 Codex. Both perform comparably for these use cases.

Token and Context Considerations

ModelMax Context WindowPractical Limit for Prompts
Claude Opus 4.6200K tokensSystem prompt + long conversations; suitable for full threat modeling sessions
GPT 5.3 Codex128K tokensSystem prompt + medium conversations
Gemini 3.1 Pro1M tokensLargest context; suitable for full-repository analysis
Grok 4.1128K tokensSystem prompt + medium conversations
GitHub CopilotVaries by planIDE context is managed by Copilot; prompt size affects available context

Prompt sizes (approximate): Most Manicode prompts fall in the 2,000-8,000 token range. Security validation prompts (High Token variants) can reach 10,000+ tokens. The AI Coding Requirements pipeline runs one stage at a time, so each stage prompt consumes roughly 3,000-6,000 tokens. Actual token counts vary by model tokenizer.

Determinism and Reproducibility

LLM outputs are inherently non-deterministic. To maximize consistency:

  • Set temperature to 0 (or the lowest available setting) for deterministic use cases like validation
  • Pin model versions in API calls (e.g., claude-opus-4-6 not claude-latest)
  • Use structured output modes (JSON mode, tool use) when available for validation prompts
  • Cache prompt-model pairs that produce good results and re-test after model updates

No combination of settings guarantees identical output across runs. Build workflows that tolerate output variance.

Cost Considerations

Prompt costs depend on the model, the prompt size, and the conversation length. General guidance:

  • Code security prompts in IDE — Cost is managed by the IDE provider's subscription (Copilot, Cursor, Windsurf)
  • API-driven workflows — Cost scales with prompt tokens (system prompt) + completion tokens (model output). High Token validation variants cost more per invocation than Low Token variants.
  • Threat modeling sessions — Multi-turn conversations accumulate context. A full 20-question interview may process 50K-100K tokens total.
  • CI/CD pipelines — Per-run costs depend on how many prompts are invoked per build. Batch validation runs are more cost-effective than per-file invocations.

Use your model provider's pricing calculator to estimate costs for your specific workflow.

Testing After Model Updates

When a model provider releases an update:

  1. Re-run your most critical prompt-model pairs against known test cases
  2. Check that security constraints are still followed (e.g., parameterized queries, input validation)
  3. Check that validation prompts still produce correct ALLOW/BLOCK decisions on your test suite
  4. If behavior has degraded, try the prompt variant for a different model as a fallback