commit dcd122ec58520026eb6e6525314b342bc3eaf1e0
Author: dtzp555-max <dtzp555-max@users.noreply.github.com>
Date:   Tue Mar 31 04:10:58 2026 +0000

    Initial commit: agent-dispatch skill for Claude Code model selection
    
    Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

diff --git a/README.md b/README.md
new file mode 100644
index 0000000..7af37e2
--- /dev/null
+++ b/README.md
@@ -0,0 +1,45 @@
+# agent-dispatch
+
+A Claude Code skill for selecting the right model (Opus/Sonnet/Haiku) when launching subagents.
+
+## What it does
+
+Before launching ANY Agent tool call, this skill evaluates the task complexity and selects the most cost-effective model:
+
+- **Opus** — Architecture decisions, domain expertise, code audits, trade-off analysis
+- **Sonnet** — Pattern-following code, known bug fixes, testing, UI work, file operations  
+- **Haiku** — Quick searches, counting, simple data extraction, status checks
+
+## Why not just use Opus for everything?
+
+Sonnet handles **70% of subagent tasks equally well** at lower cost and higher speed. This skill prevents the common pattern of defaulting to the most expensive model "just to be safe."
+
+## How it complements Superpowers
+
+The [superpowers](https://github.com/obra/superpowers) plugin includes a generic model selection guide ("cheap/standard/most capable"). This skill extends it with:
+
+- **Concrete Opus/Sonnet/Haiku mapping** for Claude Code's Agent tool
+- **Decision matrix** with specific task categories
+- **Real-world examples** from production development
+- **Split-task pattern** — e.g., Opus for research, Sonnet for implementation
+
+## Installation
+
+### As a Claude Code skill (recommended)
+
+```bash
+mkdir -p .claude/skills/agent-dispatch
+cp SKILL.md .claude/skills/agent-dispatch/
+```
+
+### In CLAUDE.md
+
+Alternatively, paste the content of `SKILL.md` into your project's `CLAUDE.md`.
+
+## Usage
+
+The skill is **not user-invocable** — it activates automatically when the agent-dispatch pattern is referenced. It guides Claude Code to select the right `model:` parameter before every `Agent` tool call.
+
+## License
+
+MIT
diff --git a/SKILL.md b/SKILL.md
new file mode 100644
index 0000000..3a8e48b
--- /dev/null
+++ b/SKILL.md
@@ -0,0 +1,61 @@
+---
+name: agent-dispatch
+description: Guide for selecting the right model (opus/sonnet/haiku) when launching subagents. Apply this before EVERY Agent tool call.
+user-invocable: false
+---
+
+# Subagent Model Selection Guide
+
+Before launching ANY Agent, evaluate the task and select the appropriate model. Default to sonnet unless the task clearly requires opus.
+
+## Decision Matrix
+
+### Use OPUS when:
+- Reading technical documentation and making architectural decisions
+- Auditing code for logical/business errors (not just syntax)
+- Designing new features with multiple valid approaches
+- Cross-referencing multiple data sources to find inconsistencies
+- Writing code that requires deep domain knowledge (e.g., ACI class hierarchy, EVPN semantics)
+- Evaluating trade-offs between approaches
+- Any task where getting it wrong is expensive to fix
+
+### Use SONNET when:
+- Writing code following an established pattern (e.g., new plugin matching existing plugin structure)
+- Fixing a known bug with clear instructions (file, line, what to change)
+- Running tests (executing curl commands, comparing results)
+- CSS/UI styling adjustments
+- File operations (git, cleanup, rename, copy)
+- Simple search-and-replace across files
+- Building frontend components with clear specs
+- Installing dependencies
+- Any task where the "what" is clear and only the "how" needs execution
+
+### Use HAIKU when:
+- Quick file searches (grep, glob)
+- Counting things (plugin count, line count)
+- Simple data extraction (parse JSON, extract field)
+- Status checks (is server running, what's the file size)
+
+## Examples from autoACI development
+
+| Task | Model | Why |
+|------|-------|-----|
+| Read Cisco ACI Multi-Site Design Guide | opus | Deep technical comprehension needed |
+| Audit 60 plugins for ACI terminology accuracy | opus | Domain expertise required to judge correctness |
+| Design EP Tunnel Trace 3-segment path | opus | Novel architecture decision |
+| Create fabric_isis plugin (follows routing_state pattern) | sonnet | Clear pattern to follow |
+| Fix contract scope context→VRF | sonnet | One-line mapping change |
+| Reorder columns in 7 plugins | sonnet | Repetitive mechanical edits |
+| Run 5-agent production test | sonnet | Execute curl, compare results |
+| Beautify CSS (fonts, colors, spacing) | sonnet | Style-only, no logic |
+| Evaluate external GitHub project | sonnet | Web search + summarize |
+| Verify topology data against APIC | sonnet | Execute queries, compare numbers |
+
+## How to apply
+
+When about to call the Agent tool:
+1. Classify the task using the matrix above
+2. Set `model: "opus"` only if the task matches OPUS criteria
+3. Default to `model: "sonnet"` for everything else
+4. Never use opus "just to be safe" — sonnet handles 70% of tasks equally well
+5. For mixed tasks (e.g., "read docs then write code"), split into two agents: opus for research, sonnet for implementation