MODEL_GUIDE.md.tmpl

skills/cowork-project-architect/templates/MODEL_GUIDE.md.tmpl

1. ---
2. type: model-guide
3. scope: project-root
4. tags:
5. - claude
6. - models
7. - reference
8. created: [YYYY-MM-DD]
9. updated: [YYYY-MM-DD]
10. aliases:
11. - Model Routing
12. ---
14. <!--
15. MODEL_GUIDE.md: Which Claude model for which task in this project.
17. Tailor to actual tasks identified in intake. Don't list every model just
18. for completeness. If the project only uses Sonnet and Haiku, only discuss those.
20. Required sections:
21. 1. Default model + reasoning
22. 2. Task-to-model table
23. 3. Decision flow (text-based)
24. 4. Concrete examples (2-3)
25. -->
27. # [PROJECT NAME]: Model Guide
29. ## Default model
31. **[Opus 4.7 / Sonnet 4.6 / Haiku 4.5]**
33. [One paragraph on why this is the default. Reference project mode (interactive/agentic), volume, and complexity profile.]
35. ## Task-to-model mapping
37. | Task | Model | Reasoning |
38. |------|-------|-----------|
39. | [Task name] | [Model] | [Why this model] |
40. | [Task name] | [Model] | [Why this model] |
41. | [Task name] | [Model] | [Why this model] |
43. ## Decision flow
45. When picking a model for a new task in this project, walk through:
47. ```
48. Is the task primarily about generating LOTS of similar outputs (>20 in a session)?
49. ├── Yes → Haiku 4.5 unless quality is critical
50. └── No → continue
52. Does the task require multi-step reasoning, novel synthesis, or careful
53. judgment across many constraints?
54. ├── Yes → Opus 4.7
55. └── No → continue
57. Is the task creative/written work, structured output, code, or routine analysis?
58. ├── Yes → Sonnet 4.6 (the default)
59. └── No → re-read the task description; if still unclear, Sonnet 4.6 and
60. escalate if needed
61. ```
63. [Add or modify branches for project-specific decision points. E.g., for a research-heavy project: "Is the task source verification with risk of hallucination? → Opus 4.7."]
65. ## Escalation triggers
67. Move up to a more capable model when:
68. - [Project-specific signal, e.g., "Output will be sent to a client without review"]
69. - [Project-specific signal, e.g., "Source verification is required"]
70. - The previous attempt at a smaller model produced a result you don't trust
72. ## Downshift triggers
74. Drop to a smaller model when:
75. - The task is genuinely repetitive and the pattern is established
76. - Latency matters more than quality margin
77. - You're burning through context on a long agentic loop
79. ## Examples
81. ### Example 1: [SCENARIO]
82. **Task:** [What's being done.]
83. **Model:** [Which one.]
84. **Why:** [One sentence.]
86. ### Example 2: [SCENARIO]
87. **Task:** [What's being done.]
88. **Model:** [Which one.]
89. **Why:** [One sentence.]
91. ### Example 3: [SCENARIO]
92. **Task:** [What's being done.]
93. **Model:** [Which one.]
94. **Why:** [One sentence.]
96. ## Cost / latency notes
98. [Only include if relevant. Skip for Cowork-only projects where the user picks the model. Include for API/batch projects with rough cost-per-task estimates and latency expectations.]
100. ## Cross-references
102. - Sub-agent model assignments: [[SUBAGENTS]]
103. - Skills with model preferences: [[SKILLS_REFERENCE]]