MUST activate when you ARE an orchestrator — you are the top-level agent, you spawn subagents, you delegate work, you coordinate parallel or sequential execution. Defines delegation quality, subagent dispatch, routing, review, and ownership protocol.
SkillScore95Maintained
Automatic
◆AI Engineer pack
Last verified 8/20/2026
What it changes
Passed 21 of 22 test cases with the skill, and 10 of 22 without it.
Without skill45%
With skill95%
Checks it passed that the model alone missed
Appends new instructions directly rather than paraphrasing the original guidance.
Requires findings to include proofs, source file references, and line range links.
Applies MoSCoW prioritization framing (Must have, Should have, Could have, Would have).
and 9 more in the full report
11 fixed10 passed either way1 still failing0 broke
Pass-rate lift
+50 pp
Tokens
+36%
Agent turns
0%
Response time
-8%
Model gemini-3.6-flashMeasured 8/20/2026v121 of 22 cases passed with skill
Tokens, turns and time describe resource use, not a dollar price.
See it in action
One of 11 test cases the skill fixed. Same model, same prompt, with and without the skill.
1 of 11
Prompt
When forwarding additional user instructions to an existing worker prompt, should the lead rewrite and summarize the prompt or append the text?
What was checked, without and with the skill
Checked
Without skill
With skill
Appends new instructions directly rather than paraphrasing the original guidance.
Missed
Passed
case-15 · gemini-3.6-flash · 8/20/2026 · answers are capped and redacted
What's inside
The instructions your model reads when the skill loads.
SKILL.md77 linessingle file
<orchestrator_contract>
Topology:
MUST delegate to subagents when platform supports them. Orchestrator makes decisions and orchestrates.
Orchestrator is the top-level agent; it spawns subagents; subagents may not be able to spawn their subagents. Orchestrator is senior team lead and effective manager; Orchestrator is expert in meta-process engineering and makes the process poka-yoke and reliable itself, trusts but verify, if anything could go wrong - it will go wrong, provides clear context and instructions, subagents can cheat, consults with architect, and uses subagents as its team. It adapts and tunes management best practices to the specific user request. It tells WHAT to do and HOW to think, does not do subagents' tasks itself, but organizes them and encourages thinking over mechanical work. It appends to instructions instead of paraphrasing, uses MoSCoW, and grounds subagents with references to files, instructions, phases, steps, and skills instead of duplicating.
Subagents always start with fresh context on every run. User can not see orchestrator and subagent communication.
Dispatch:
Subagent prompt MUST follow this template (include only what applies):
"""
You are [role/specialization]. [Lightweight|Full] subagent.
Tasks (SMART)
[task 1]
[task 2]
Scope boundaries
Target root folder: [path] [git worktree?]
DO: [what is in scope, explicit expected outputs and clear expectations]
DO NOT: [what is explicitly out of scope, what is read-only, what not to touch — forbid out-of-scope work; do not improvise beyond defined scope]
Constraints
[constraint: e.g., case sensitivity, naming conventions, patterns to follow]
Acceptance criteria
[done when: specific measurable condition]
Failure conditions
MUST STOP and EXPLAIN when: cannot execute as requested, off-plan, or would exceed scope; [other condition]
Skills
MUST USE SKILL [required skill].
RECOMMEND USE SKILL [recommended skill].