delegate-kit: 8 Claude Code subagents + playbook

Stop burning expensive context on long Claude Code sessions. Large objectives often lead to massive context accumulation, driving up token costs and slowing down response times. This skill provides a specialized orchestration layer that delegates work to a fleet of eight tiered subagents, each optimized for specific reasoning depths and context costs.

What it does

• Multi-agent delegation routes work to eight specific subagent types including thinkers, researchers, and executers.
• Cost optimization: measured across 1,233 real Claude Code sessions, running as the director cut cost 6-10x per unit of work on objectives past ~35 tool calls, because the expensive model never accumulates a context. Below that line it costs more - full numbers, both directions, are in reference/measurements.md.
• Automated partitioning splits wide, multi-file changes into independent work units and parallel execution clusters.
• Web-enabled research provides a dedicated agent for fetching external documentation and API specs without repo access overhead.
• Defect hunting fans out cheap, disposable agents to sweep a listed set of files or a diff for real defects (never style).

How it works

1. Setup installs the agent definitions and Python-based partitioning tools to the local ~/.claude directory.
2. Orchestration initiates a session using the director agent to manage the high-level objective and delegation logic.
3. Partitioning uses the partitioner to map out complex tasks into a schedule.json before any implementation begins.
4. Execution spawns targeted workers (thinker, executer, simple-tasks) to perform isolated units of work and report findings by reference.

Frameworks & tools

The kit requires Claude Code and Python 3. It utilizes built-in shell tools like grep, bash, and sed, with optional enhanced support for Serena and graphify MCP servers.

Why this beats prompting it yourself

Writing manual prompts for subagents requires constant context management to prevent token bloat. This skill provides a formal protocol for "relay by reference," moving payloads through the filesystem instead of the chat history.

Use cases

• Refactoring large subsystems where a single agent would exceed context limits.
• Performing parallel bug sweeps across dozens of files simultaneously.
• Implementing features that require deep external API research before coding.
• Automating mechanical chores like git management and multi-hop builds.

Known limitations

Claude Code only: the director and the per-agent model pins rely on Claude Code subagents. Requires a manual installation step and a new session start to load agent definitions. Subagents are instructed not to run destructive git commands; this package installs no hooks, so that rule is enforced by their instructions only.