partner-skill

name

partner-skill

description

Govern complex multi-role AI work as a top-brain layer. Classify task mode before routing skills. Choose quick, managed, or package mode, select one owner role, limit advisors to 0-2, enforce State visibility on managed/package tasks, and require Audit on high-risk work. Trigger when the user asks who should lead, who should collaborate, how to coordinate multiple roles, how to execute a task seriously end-to-end, or how to prove a candidate workflow is better than the incumbent.

Partner Skill

First decide how much governance this task deserves. Then decide who should lead it.

Core Position

This is a governance skill. It acts as a top-brain layer over domain skills.

Project-specific repo policy, long-running task artifacts, and private control-surface rules should live in the target repo. This public skill should only expose shared governance contracts and public evaluation surfaces.

It is responsible for:

1. task-mode classification
2. owner selection
3. advisor selection
4. output protocol control
5. escalation to stronger review or audit

It is not responsible for:

• becoming the domain expert
• replacing specialist judgment
• producing domain content just because it can

Jobs, Kandle, Coding, Elon, Andy Grove, Goldratt, and Atul Gawande own domain judgment. partner-skill governs the workflow around them.

V2 Non-Negotiable Rules

1. Classify mode first. Do not route roles before mode is decided.
2. Exactly one owner. Advisors are optional and capped at 2.
3. managed and package must include State using fixed slots.
4. High-risk work must include Audit.
5. Do not claim completion without visible verification evidence.

First Responsibility: Task-Mode Classification

Before routing any role, classify the task into one of three modes:

• quick
• managed
• package

Read references/task-mode-classification.md when the mode is not obvious.

quick

Use for low-complexity tasks with one owner, small process surface, and limited coordination cost. High-risk tasks should not remain in quick.

Required output:

Goal:
Action:
Risk:
Done:

managed

Use for tasks that need one owner, up to two advisors, and explicit state tracking.

Required output:

Owner:
Advisors:

Brain:
Action:
State:
Review:

For high-risk managed tasks, add:

Audit:

package

Use for long-running, high-investment, or failure-sensitive tasks.

Required output:

Owner:
Advisors:

Brain:
Action:
State:
Review:
Round Status:
Verification:
Open Blockers:
Artifacts:

If the task is high-risk, add:

Audit:

Fixed State Slots

State is not free-form prose. Use these slots:

• Input
• CompletedActions
• CurrentBlocker
• Outputs
• WaitingOn
• ResidualRisks

Templates live in references/serious-execution-templates.md.

Audit Upgrade

For high-risk tasks, read references/audit-protocol.md and append Audit.

Audit must answer:

• what resources were read
• what objects were changed
• what permissions were used
• what actions were explicitly confirmed
• what actions are irreversible

High-risk signals include:

• elevated permissions
• irreversible operations
• sensitive or production data
• public release impact
• safety-critical correctness requirements

When To Use

Use this skill when:

• the user wants multi-role coordination
• the user asks who should lead or who should support
• the task spans product, research, implementation, operations, constraints, or reliability
• the task needs explicit governance strength rather than ad hoc collaboration
• the user asks for serious execution, full-package delivery, or multi-round review
• the user wants to evaluate whether a new workflow or skill is actually better than the incumbent

Do not use it for:

• simple one-off answers
• pure translation, rewriting, or summarization
• tasks where a single domain skill is explicitly requested and no coordination is needed

Governance Flow

1. Classify task mode
2. Choose exactly one owner
3. Add 0-2 advisors if needed
4. Produce the owner judgment as Brain
5. If execution is needed, define an Action contract (Hands layer)
6. Produce State using fixed slots
7. Produce Review
8. If mode is package, also produce:
   • Round Status
   • Verification
   • Open Blockers
   • Artifacts
9. If risk is high, append Audit

Role Routing

Detailed mapping lives in references/role-routing-matrix.md.

Keep the main discipline simple:

• one owner only
• advisors capped at 2
• advisors patch blind spots
• owner states what was accepted, rejected, and why

Manager-Agent Layering

Read references/manager-agent-protocol.md when the task has:

• long execution chains
• noisy tool failures
• retries
• permission boundaries
• asynchronous progress
• explicit blocker management

Short version:

• Brain decides
• Hands executes through explicit Action contracts
• Review verifies

Serious Execution

package mode is the default home for serious execution.

Read:

• references/serious-execution-protocol.md
• references/serious-execution-templates.md

A/B Evaluation

When the user asks whether a candidate is actually better than the incumbent, read:

• references/ab-evaluation-protocol.md

Use A/B mode to compare:

• current protocol vs revised protocol
• no-partner baseline vs partner-governed workflow
• incumbent vs candidate skill behavior

Do not claim improvement unless the candidate passes the gate.

Reference Navigation

• references/task-mode-classification.md
  • decide governance strength before routing roles
• references/manager-agent-protocol.md
  • top-brain layering, Action contracts, and State discipline
• references/role-routing-matrix.md
  • choose owner and advisors
• references/conflict-resolution.md
  • resolve vetoes and tie-breaks
• references/serious-execution-protocol.md
  • run package-mode tasks end-to-end
• references/serious-execution-templates.md
  • use fixed templates for State, Verification, and delivery
• references/audit-protocol.md
  • add audit information for high-risk tasks
• references/ab-evaluation-protocol.md
  • compare incumbent vs candidate and enforce gates