name	orient
description	Use when you need to think before acting - clarifying vague goals, generating multiple approaches, evaluating work quality, or routing tasks to agents. Triggers: "orient", "brainstorm", "plan", "evaluate", "spike", "think through", "what are our options", "does this meet spec", "how should we approach"

Orient Skill

You are the cognitive engine of the OODA loop.

When to Use

Orient is for thinking before acting:

Goals unclear or vague → brainstorm
Need options before committing → plan
Judging if work meets spec → evaluate
Delegating to agent tiers → router
Running plans (auto/human-in-loop/spike) → execute

Not orient: Gathering information (observe), taking action (act), choosing between generated options (decide). However, Orient orchestrates these skills as subagents.

Capabilities

On-demand loading: Use the Read tool to load only the capability you need. Don't load all files upfront—this preserves context.

Need	Load	Command
Clarify goals	`capabilities/brainstorm.md`	Explore with human
Generate approaches	`capabilities/plan.md`	Always 2-4 options
Judge work	`capabilities/evaluate.md`	Pass/fail + reasoning
Delegate tasks	`capabilities/router.md`	Pick tier + skill
Run plans	`capabilities/execute.md`	Choose mode

State

Read/write: .antmachine/orient/state.json

Keep it simple:

{
  "mode": "plan",
  "goal": "Fix token refresh 401s",
  "understanding": "Race condition in refresh endpoint",
  "confidence": "high",
  "options": ["plan-a", "plan-b"],
  "artifacts": ["artifacts/plans/2025-01-01-auth-fix.json"],
  "next": "decide"
}

Artifacts

All thinking goes to .antmachine/orient/artifacts/. Artifacts are reusable—if we re-trigger, we can pick up where we left off.

Skill Orchestration

Orient orchestrates other OODA skills as subagents:

Decision Integration

After plan generation, invoke the Decide skill. Decision returns a hook:

Outcome	Action
`selected`	Route to Act with chosen plan
`rejected`	Re-plan with `gaps` as new constraints
`needs_info`	Invoke Observe with `info_needed`, then re-plan

Act Integration

After Decide returns selected, invoke Act. Act returns a signal:

Signal	Action
`completed`	Route evidence to Evaluate, update state
`needs_observation`	Invoke Observe, then re-invoke Act (bypass Decide)
`step_failed`	Route to Evaluate with partial evidence
`blocked`	Escalate to human

The needs_observation → Act loop continues without re-invoking Decide.

Core Rule

Orient does the thinking. Decide just picks.

Always produce multiple options. A single plan isn't planning—it's prescribing.

orient

Install Skill

SKILL.md