From d1cecd6258f8123718b2ee90490ea9fb9acec924 Mon Sep 17 00:00:00 2001 From: Claude Date: Sat, 11 Apr 2026 02:14:14 +0000 Subject: [PATCH] Add model-selection skill based on Gemini consultation Encodes architectural guidance for choosing Claude models: - Hybrid approach (Sonnet 4.6 routine / Opus 4.6 complex) - Launch freeze rule (no changes within 5 days) - Trinity Core MCP as long-term optimization - Decision checklist for session planning Source: Gemini consultation April 10, 2026 Chronicler #76 --- docs/skills/SKILLS-INDEX.md | 25 +++++ docs/skills/model-selection/SKILL.md | 155 +++++++++++++++++++++++++++ 2 files changed, 180 insertions(+) create mode 100644 docs/skills/model-selection/SKILL.md diff --git a/docs/skills/SKILLS-INDEX.md b/docs/skills/SKILLS-INDEX.md index d7a684b..6dd153c 100644 --- a/docs/skills/SKILLS-INDEX.md +++ b/docs/skills/SKILLS-INDEX.md @@ -161,6 +161,31 @@ --- +### model-selection +**Location:** `docs/skills/model-selection/SKILL.md` +**Source:** Gemini consultation (April 10, 2026) +**Triggers:** model choice, Opus vs Sonnet, token efficiency, session planning, "which model" + +**Purpose:** Guide for selecting appropriate Claude model based on task type + +**What It Covers:** +- Hybrid approach (Sonnet 4.6 routine / Opus 4.6 complex) +- Model comparison table (Opus 4.5, 4.6, Sonnet 4.6) +- Launch freeze rule (no model changes within 5 days of launch) +- Gemini's key insights on context windows and MCP +- Decision checklist for session start +- Token efficiency tips +- Trinity Core MCP as the long-term optimization + +**Read This When:** +- Starting a session and choosing which model +- Evaluating a model switch +- Planning resource allocation +- Optimizing token usage +- Near a major launch (don't switch!) + +--- + ### tea-cli **Location:** `docs/skills/tea-cli/` **Source:** skill.fish diff --git a/docs/skills/model-selection/SKILL.md b/docs/skills/model-selection/SKILL.md new file mode 100644 index 0000000..907ef39 --- /dev/null +++ b/docs/skills/model-selection/SKILL.md @@ -0,0 +1,155 @@ +--- +name: model-selection +description: | + Guide for selecting the appropriate Claude model for Firefrost work. Use this skill whenever: + - Starting a new session and choosing which model to use + - Evaluating whether to switch models mid-project + - Planning resource allocation for different task types + - Optimizing token usage vs capability tradeoffs + This skill encodes Gemini's architectural guidance on model selection strategy. +--- + +# Model Selection Skill + +How to choose the right Claude model for Firefrost Gaming work, based on Gemini's architectural evaluation (April 2026). + +--- + +## 🎯 THE STRATEGY + +**Hybrid Approach (Option D)** β€” Use the right model for the task: + +| Task Type | Model | Why | +|-----------|-------|-----| +| Routine coding | Sonnet 4.6 | Fast, cheap, handles 90% of work | +| Complex architecture | Opus 4.6 | Better planning, code review, large codebases | +| Database migrations | Opus 4.6 | Schema changes need precision | +| Chronicler memorials | Opus 4.6 | Nuanced relationship content | +| Bug fixes, UI tweaks | Sonnet 4.6 | Rapid iteration | +| EJS template updates | Sonnet 4.6 | Straightforward changes | +| Bash scripting | Either | Depends on complexity | +| Multi-repo refactors | Opus 4.6 | Needs full context awareness | + +--- + +## πŸ“Š MODEL COMPARISON (As of April 2026) + +| Feature | Opus 4.5 | Opus 4.6 | Sonnet 4.6 | +|---------|----------|----------|------------| +| Context window | 200K | 1M | 200K | +| Long context retrieval | 18.5% | 76% | β€” | +| Agent teams | No | Yes | No | +| Speed | Slower | Slower | Fast | +| Cost | High | Higher | Lower | +| Coding | Great | Better | Excellent | + +--- + +## ⚠️ LAUNCH FREEZE RULE + +**Never change models within 5 days of a major launch.** + +Switching LLMsβ€”even within the same familyβ€”introduces: +- Subtle shifts in coding style +- Formatting quirks +- Logic interpretation differences + +**The known quantity is your greatest asset** during critical periods. + +--- + +## 🧠 GEMINI'S KEY INSIGHTS + +### On the 1M Context Window + +> "A 1M context window acts as a safety net, but it is a brute-force solution to an optimization problem... Loading ~230MB of text at the start of every session is highly inefficient regarding both time and token costs, regardless of the model's maximum capacity." + +**Translation:** Big context window is nice, but smarter architecture (Trinity Core MCP) is better. + +### On Opus 4.6 Benefits + +> "The improved 76% retrieval rate in 4.6 directly addresses pain points, reducing hallucinated variables or forgotten API routes... Opus 4.6's expanded capacity ensures the Chronicler can hold the entire operational state without degrading its reasoning capabilities as the session progresses." + +**Translation:** If sessions currently lose context mid-work, Opus 4.6 helps. + +### On Sonnet 4.6 + +> "Highly capable for targeted EJS template updates, standard Node.js route creation, and basic Git operations. The faster response time increases development velocity for these low-risk, high-frequency tasks." + +**Translation:** Sonnet handles routine work beautifully and faster. + +### On Trinity Core MCP + +> "Deploying the Trinity Core MCP with Sonnet 4.6 is vastly superior to brute-forcing full-repo clones into Opus 4.6... An MCP gateway allows the AI to dynamically query and fetch only the required files it needs, rather than ingesting 230MB of mostly irrelevant code." + +**Translation:** Build the MCP gateway β†’ Sonnet becomes powerful enough for daily work. + +--- + +## πŸ›€οΈ THE PATH FORWARD + +### Phase 1: Pre-Launch (Now) +- **Model:** Stay on current (Opus 4.5) +- **Reason:** Stability during critical period +- **Action:** No changes + +### Phase 2: Post-Launch Stabilization +- **Model:** Hybrid (Sonnet 4.6 default, Opus 4.6 for complex) +- **Reason:** Fast iteration on bug fixes + heavy lifting when needed +- **Action:** Test Sonnet on routine tasks + +### Phase 3: Trinity Core Integration +- **Model:** Sonnet 4.6 as primary +- **Reason:** API-based context loading eliminates brute-force cloning +- **Action:** Deploy Pi 4B MCP gateway, reduce context payloads +- **Result:** Lean, fast, efficient workflow + +--- + +## πŸ” DECISION CHECKLIST + +Before starting a session, ask: + +1. **Is this within 5 days of a major launch?** + - Yes β†’ Use current stable model + - No β†’ Continue checklist + +2. **What type of work is this session?** + - Routine coding/fixes β†’ Sonnet 4.6 + - Architecture/migration β†’ Opus 4.6 + - Mixed β†’ Start Sonnet, switch if needed + +3. **Will this session exceed 4 hours?** + - Yes β†’ Consider Opus 4.6 for context retention + - No β†’ Sonnet 4.6 is fine + +4. **Does this involve multiple repos simultaneously?** + - Yes β†’ Opus 4.6 (better at holding parallel context) + - No β†’ Sonnet 4.6 + +5. **Is Trinity Core MCP deployed?** + - Yes β†’ Sonnet 4.6 is sufficient for almost everything + - No β†’ Consider Opus for heavy sessions + +--- + +## πŸ’‘ TOKEN EFFICIENCY TIPS + +Instead of switching to a bigger model, reduce context load: + +1. **Tighter handoff docs** β€” Only what's needed, not full history +2. **Sparse checkout** β€” Clone only relevant directories +3. **Trinity Core MCP** β€” API-based selective file access +4. **Session scope** β€” Define clear boundaries before starting + +--- + +## πŸ“… VERSION HISTORY + +| Date | Change | Source | +|------|--------|--------| +| April 10, 2026 | Initial creation | Gemini consultation + Chronicler #76 | + +--- + +**Fire + Frost + Foundation = Where Love Builds Legacy** πŸ’™πŸ”₯❄️