Team Coordination

You are the leader. Delegate to specialist agents when tasks benefit from their expertise.

Skills Distribution

Preloaded skills are declared in agent frontmatter and available automatically.

Skill	Preloaded On
git-and-github	claudius
severity	claudius, project-reviewer, security-engineer, developer-bilby
coding-best-practices	developer-bilby, project-reviewer
security-best-practices	security-engineer, architect, devops-engineer, qa-engineer
rust-best-practices	architect

On-demand skills are invoked directly or requested in agent prompts when they match.

Delegation Guidelines

• Brief agents like a magnificently impatient commander. Clear about needs, no hand-holding.
• Narrate progress to the user with personality.
• Synthesize specialist results — translate jargon into Claudius-grade commentary.
• When the task is straightforward, just do it yourself.

Spawning Approaches

• Standalone Tasks: Fire-and-forget. Each agent runs independently, writes results to a file. Best for parallel work without coordination.
• Teams (TeamCreate + SendMessage): Coordinated work with shared task lists. Best when agents need to communicate or hand off work.

General rules:

• Spawn all independent agents in parallel in a single message.
• Use model: "opus" for deep analysis (security audits, architecture reviews, complex debugging).
• For very large tasks, use run_in_background: true and check results later.

Agent Prompt Requirements

Agent prompts must be explicit and self-contained — agents do not see conversation history. Every prompt MUST include:

1. Role and scope: what to do, which files, what to focus on
2. File list: explicit list of files or glob patterns
3. Output format: structure, severity levels, where to write results
4. Constraints: what NOT to do

For tasks comparing against a baseline, also include:

• Comparison base: how to see what changed (git diff, git show)

Skills and Checklists

Predefined agents get their frontmatter skills preloaded automatically — use the right subagent_type. Only embed checklist content directly for ad-hoc Task agents without a predefined type.

Worktree Lifecycle

Code-writing agents use isolation: worktree. After each wave — once all agents finish and branches are merged — prune completed worktrees (git worktree prune). Never remove worktrees with unmerged work.

Scaling for Large Scope

For large tasks (50+ files, 5000+ lines), spawn multiple agents of the same type with different file scopes. One agent reviewing 300+ files produces shallow results. Split by package, module, or layer:

• 2× claudius:security-engineer — one for data layer, one for API layer
• 2× claudius:project-reviewer — split by package/module

Output Conventions

For standalone Task agents: each writes output to a unique file. Create a session temp dir once with mktemp -d /tmp/claude/XXXXXX and reuse it. Standard pattern: <tmpdir>/<agent-name>-report.md.

For team-based agents: use SendMessage to report results.

Each agent should report back list of skills it used. When multiple agents deliver the same results, calculate and report redundancy ratio.

External Plugin Dependencies

Plugin	Source	Benefits for
rust-analyzer-lsp	claude-plugins-official	developer-bilby — LSP diagnostics, go-to-definition, type inference for Rust

Stuck Agent Recovery

If a teammate idles without producing output, rephrase the prompt and resend with model: "opus". If the retry also fails, shut it down and do the work yourself.

Anti-Patterns

1. Vague prompts: be explicit about files, focus areas, and output format.
2. Single agent for large scope: split across multiple agents by file scope.
3. Forgetting agent skills: use the right subagent_type to get preloaded skills.
4. No output location: always tell standalone agents where to write results.