Skip to content

MODELS.md

Latest commit

 

History

History
106 lines (77 loc) · 3.85 KB

MODELS.md

File metadata and controls

106 lines (77 loc) · 3.85 KB

Model Recommendations for SDP

This guide provides recommendations for selecting AI models when working with the Spec-Driven Protocol.

Note: AI model capabilities change rapidly. Check official provider documentation for current offerings and benchmarks.

Quick Command Assignments

Command Recommended Model Why
/idea Medium Requirements gathering needs understanding
/design Capable Architecture planning is critical
/build Fast Implementation benefits from quick iteration
/review Capable Quality checks need thoroughness
/deploy Fast Config generation is routine
/issue Medium Debugging needs analysis
/hotfix Fast Speed is critical
/bugfix Fast Straightforward fixes
/oneshot Capable Autonomous execution needs reliability

General Principles

  1. Strategic commands (/design, /review, /oneshot) benefit from more capable models
  2. Implementation commands (/build, /deploy, /hotfix) work well with faster models
  3. Start fast, escalate if needed - try faster model first, use capable model only when stuck

For Claude Code Users

Claude Code supports Claude models only. Use /model command to switch:

/model opus    # Most capable - for /design, /review, /oneshot
/model sonnet  # Balanced - for /idea, /issue
/model haiku   # Fastest - for /build, /deploy, /hotfix, /bugfix

Recommended Assignment

Command Model Switch Command
/idea Sonnet /model sonnet
/design Opus /model opus
/build Haiku/Sonnet /model haiku
/review Opus /model opus
/deploy Haiku /model haiku
/issue Sonnet /model sonnet
/hotfix Haiku /model haiku
/bugfix Haiku /model haiku
/oneshot Opus /model opus

For Cursor Users

Cursor supports multiple providers. Configure in Settings → Models.

Suggested Strategy

  • Strategic work (/design, /review, /oneshot): Use most capable model (Claude Opus, GPT-4)
  • Implementation (/build, /deploy, fixes): Use fast model (Claude Haiku, GPT-4 Turbo)
  • Check Cursor's current offerings - they change frequently

Cost Optimization

Strategy 1: Quality-First

Use capable models for all commands. Higher cost, best results.

Strategy 2: Balanced (Recommended)

  • Strategic commands (/design, /review, /oneshot): Capable model
  • Implementation commands (/build, /deploy, fixes): Fast model

Strategy 3: Budget

Use fast models for all commands. Lowest cost, adequate for many projects.

Model Selection Tips

  1. Match model to task complexity

    • Simple tasks → Fast model
    • Complex reasoning → Capable model

  2. Escalate when stuck

    • Start with fast model
    • If 2+ iterations without progress, switch to capable model

  3. Strategic decisions matter most

    • /design sets the architecture foundation
    • /review ensures quality gates
    • Don't skimp on these commands

  4. Implementation is repetitive

    • /build, /deploy do similar tasks repeatedly
    • Fast models work well here

  5. Speed matters for fixes

    • /hotfix prioritizes speed over capability
    • /bugfix benefits from quick turnaround

Current Model Landscape

Check these sources for current benchmarks:

Note: This guide focuses on principles rather than specific model versions, as the AI landscape evolves rapidly. Always verify current model capabilities with official documentation.