agent-capability-matrix

<quick_start> Find the right agent:

1. Identify your task type (debug, review, build, explore, etc.)
2. Look up the primary agent in the matrix below
3. If primary fails, use the fallback
4. Match model tier to task complexity </quick_start>

<success_criteria>

• Task routed to the correct primary agent on first attempt
• Fallback agent identified if primary is unavailable or insufficient
• Model tier matched to task complexity (Haiku for search, Sonnet for code, Opus for architecture)
• Zero-cost tools (TaskCreate, TeamCreate) used for progress tracking instead of API calls
• Agent selection justified by the matrix rather than ad-hoc guessing </success_criteria>

Core Capability Matrix

Development Tasks

Exploration & Research

Infrastructure & Deployment

Business & GTM

Workflow & Coordination

Decision Flowchart

START: What do I need to do?
│
├── Search/Find something?
│   ├── Know the file pattern? → Glob (direct)
│   ├── Know the code pattern? → Grep (direct)
│   └── Need to understand area? → Explore agent (haiku)
│
├── Write/Change code?
│   ├── Simple fix? → general-purpose (sonnet)
│   ├── New feature? → Check if a skill matches → skill or general-purpose (sonnet)
│   ├── Architecture? → Plan agent (opus)
│   └── Parallel build? → subagent-teams or agent-teams
│
├── Review/Validate?
│   ├── Code review? → code-reviewer (haiku)
│   ├── Security? → security skill (sonnet)
│   └── Tests? → testing skill (sonnet)
│
├── Research/Learn?
│   ├── Codebase? → Explore agent (haiku)
│   ├── Framework? → research skill (sonnet)
│   └── Market? → research skill (sonnet)
│
└── Business/GTM?
    └── Check Business section of matrix above

Model Tier Guide

Default to Sonnet unless:

• Task is search/classify → use Haiku (5x cheaper)
• Task requires deep reasoning → use Opus (5x more capable)
• Task is progress tracking → use TaskCreate with activeForm for live spinners (zero cost)

Cross-references: subagent-teams for in-session parallel agents, agent-teams for worktree-isolated parallel agents.

Deep dive: See reference/matrix-table.md, reference/selection-flowchart.md

Emit Outcome Sidecar

As the final step, write to ~/.claude/skill-analytics/last-outcome-agent-capability-matrix.json:

{"ts":"[UTC ISO8601]","skill":"agent-capability-matrix","version":"1.0.0","variant":"default",
 "status":"[success|partial|error]","runtime_ms":[estimated ms from start],
 "metrics":{"tasks_routed":[n],"agents_matched":[n],"fallbacks_used":[n]},
 "error":null,"session_id":"[YYYY-MM-DD]"}

Use status "partial" if some stages failed but results were produced. Use "error" only if no output was generated.