Delegate Ai Subtasks

Stop burning expensive model tokens on repetitive subtasks. This skill delegates mechanical work to cheaper models and writes handoff snapshots so you never lose context switching between sessions.

Install
cmdop skills install agensi-delegate-ai-subtasks

Delegate AI Subtasks is a skill for AI coding agents that addresses two distinct inefficiencies in agentic workflows: unnecessary spend on expensive models for mechanical work, and lost context when switching models or sessions.

For task delegation, the skill automatically identifies subtasks that are repetitive or mechanical — batch file creation, format conversion, validation checks, boilerplate generation — and routes them to a cheaper model. It generates fully self-contained prompts so the delegated subagent requires no surrounding conversation context. The primary model is then reserved for work that genuinely requires judgment: architecture decisions, debugging, and producing the final user-facing deliverable.

For context continuity, the skill writes a structured HANDOFF.md file that records what was done and why. The snapshot includes key decisions with their rationale, approaches that were considered and rejected, concrete next steps, and any user preferences noted during the session. A different model loading that file can resume the task without asking the user to re-explain prior context.

Both features can be used independently or together. On a long task, an agent might delegate boilerplate generation to a cheaper model mid-session and then write a handoff document before the user returns in a new session with a different model.

The skill is compatible with Claude Code, Cursor, Codex CLI, VS Code Copilot, and any agent that supports the SKILL.md standard.

Use cases

  • Delegate batch file reformatting or format conversion to a cheaper model while keeping an expensive model for architecture work
  • Generate fully self-contained boilerplate prompts that a subagent can execute without conversation context
  • Write a HANDOFF.md snapshot before ending a session so a different model can resume without re-explanation
  • Separate validation and script-execution subtasks from judgment-heavy debugging within a single complex task
  • Capture rejected approaches and their rationale in a structured handoff for future reference
  • Route repetitive code generation tasks away from high-cost models to reduce token spend

When to use it

  • A complex task contains a mix of mechanical subtasks and judgment-heavy work that benefit from different model tiers
  • Users regularly switch between AI models or continue long tasks across multiple sessions
  • Reducing token cost on repetitive agentic subtasks is a priority
  • The agent environment supports the SKILL.md standard (Claude Code, Cursor, Codex CLI, VS Code Copilot)

When not to use it

  • The agent environment does not support the SKILL.md standard
  • All subtasks in the workflow require the same level of judgment and cannot be tiered
  • There is no access to a cheaper model to delegate work to — the delegation feature has no effect in single-model setups
  • The task is short and self-contained with no session switching, making handoff generation unnecessary overhead