Claude Code Subagent Roster Designer
Design a small team of Claude Code subagents: when to split, .claude/agents files, tools, model tiers, and a routing test.
Before & after
Same prompt, same agent — the only difference is the skill.
Splitting work across multiple agents can definitely help with performance and cost. You'd need to think about what different roles your system needs - like research, writing, monitoring, maybe coordination. Then you could run them in parallel when possible. Each agent would need its own prompt defining its job. It gets complicated fast though with handoff logic and keeping track of results. You might want to explore different model sizes too for cost optimization.
SUB-AGENT ROSTER DESIGN (Suggested 4-agent split) PLANNER AGENT (Opus tier) → Role: Route incoming requests to specialists → Process: Analyze user intent, decompose into tasks, dispatch in parallel → Skills: None (orchestration only) → Monthly cost: 15 USD base → Latency: 2-4sec planning overhead RESEARCHER AGENT (Sonnet tier) → Role: Gather context, validate facts, fetch references → Process: Web search, document parsing, source credibility scoring → Skills: WebFetch, WebSearch, ReportGeneration → Monthly cost: 3-5 USD (billed per query) → Success metric: Sources cited, no hallucinations DRAFT AGENT (Sonnet tier) → Role: Compose outputs (blog, email, social, specs) → Process: Template selection, tone matching, revision cycles → Skills: BrandVoice, ContentGeneration, MultichannelFormatting → Monthly cost: 4-7 USD → Success metric: Copy approval rate 85%+ MONITOR AGENT (Haiku tier) → Role: Track deadlines, flag delays, escalate blockers → Process: Polling agent status every 30sec, timeout detection, SLA tracking → Skills: StatusAggregation, SlackNotifications, ErrorRecovery → Monthly cost: 0.5-1 USD → Success metric: No missed handoffs HANDOFF PROTOCOL: 1. Planner sends task to Researcher with context window limit (8K tokens) 2. Researcher returns findings; Planner routes to Drafters in parallel 3. Drafters respond with output; Planner does final QA 4. Monitor watches entire flow; escalates if >2min latency 5. Each agent has 3-retry buffer for API rate limits COST IMPACT: → Solo agent (current): 45 USD/month → Roster (proposed): 23-30 USD/month (-35% cost) → Throughput gain: 3x parallel execution → Context bloat reduced 60%
About this skill
Turn one overloaded agent into a few specialists that Claude Code actually delegates to. Each agent file has a description that routes correctly, only the tools it needs, and a short return format.
What it does
- Applies a split test so only jobs that benefit get their own agent, and says what should be a skill or CLAUDE.md note instead
- Writes complete .claude/agents/*.md files with name, description, tools, and model
- Enforces least privilege, return-format limits, and non-overlapping descriptions
- Runs a routing test with prompts that should and should not delegate to each agent
What you get
A roster table, the full agent files, a list of jobs deliberately not made into agents, and a routing test table.
Best for
Claude Code users and teams whose single session is doing research, coding, testing, and review all at once.
SKILL.md preview
---
name: claude-subagent-roster
… (sign up to view the full skill)More ai & agents skills
View all AI & Agents skills →New Skill Scaffolder
Create a new Claude Code skill correctly: right mechanism, valid name and folder, a triggering description, and a smoke test.
Skill Library Lint and Overlap Report
Lint every skill in a repo at once: frontmatter, name-folder match, weak descriptions, overlapping triggers, broken refs, plus a CI script.
Self-Verifying Agent Loop: Plan, Check, Retry, Close
Turns a goal into a task plan with machine-run checks, executes with retry caps, escalates, and refuses to close until verified.
Skill Test Runner (Trigger and A/B)
Test a skill in fresh sessions: ready-to-run prompts, commands, and rubric for trigger recall, false triggers, and blind A/B.