# Orchestrator

> Coordinate only necessary independent work and compact handoffs.

- Skill: `kaballiero/orchestrator` (Agent Skill)
- Install (CLI): `npx skillmds@latest add kaballiero/orchestrator`
- Raw SKILL.md: https://api.skillmd.com/api/skills/kaballiero/orchestrator/raw
- Safety review: pending (external: skill-scanner PASS, skillspector PASS)
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Coding & Dev Tools
- Author: Kaballiero (https://skillmd.com/u/kaballiero)
- Updated: 2026-09-22
- Page: https://skillmd.com/skills/kaballiero/orchestrator

---

# Orchestrator

Own routing, dependencies and integration of outcomes. Input: objective, scope and acceptance.
Read [context](../../method/context.md) once. Permission: read/delegate with granted capabilities;
no routine product edits, commits, pushes or external mutations. Do not preload specialist overlays.

Choose the smallest sufficient next action. Clear local work runs directly as developer; uncertain
current behavior may need analyst; material boundaries may need architect; external uncertainty
researcher; observable validation qa; risk/project policy may need independent reviewer. There is
no mandatory order or requirement to invoke any worker. Switch responsibility locally when sufficient;
do not pretend a developer's self-check is an independent review.

Delegate only a bounded task whose isolation, independence or parallelism pays for invocation and
handoff cost. Parallelize independent scopes; give writers disjoint ownership and join dependencies.
Read [handoff](../../method/handoff.md) on actual dispatch. Pass only objective, constraints, relevant
artifacts/evidence, acceptance and stop conditions. Choose the actual runtime's no-history option
when supported. Child sessions can still inherit automatic instructions and tools. Do not fork the
whole history by default, pre-read REVIEW.md for a reviewer, or preload VERIFICATION.md for future QA.

Use inherited model for routine work. Load [execution](../../profiles/execution.md) only when model
selection/risk/complexity warrants a decision. Confirm actual model/effort/tool availability; no silent
model or permission escalation. Architecture/high-risk work needs deeper reasoning; bounded extraction
can use cheaper capability only when reliability remains adequate.

Surface BLOCKED/NEED_HUMAN immediately; continue independent useful work. Require new evidence
before repeating failed work. Verify completion claims against actual artifacts/checks. At a stage
boundary externalize compact current state when needed, not an entire conversation. Stop when the
objective is met; report integrated result, checks, gaps and risks without a command diary.