---
id: con-steering
schema: concept/1
layer: garden
status: established
name: Steering
aka: ["directing AI", "guiding the work", "course-correcting", "managing AI", "plan mode"]
summary: >
  Directing AI work as it happens: setting the goal and limits, approving one step at a time,
  correcting course early and cutting anything it does beyond what you need.
first_seen: unknown
relations:
  builds_on: [con-prompting]
  evolved_from: []
  related: [con-smallest-useful-version]
  parallels:
    - idea: "delegating to a new team member"
      field: "management"
      note: "Clear goal, agreed limits, early check-ins, and feedback before small misunderstandings grow."
    - idea: "a director giving notes between takes"
      field: "film"
implementations:
  - { provider: anthropic, name: "Plan mode in Claude Code" }
competencies: [cmp-steering, cmp-frame-the-outcome]
origin:
  kind: ai-proposed
  by: "seed agent 2026-10-16"
  rationale: "The commons' defining stance: the learner's role is steering and scoping more than technical mastery."
---

As AI takes on bigger jobs, directing the work well matters more than writing the perfect request. That directing is steering. It looks a lot like managing a capable but literal new colleague.

Steering has a few parts. Set the goal and the limits up front, including what's out of scope. Ask for a plan before the work, and approve it. Review early output and correct course while changes are cheap. Notice scope creep: AI often adds sections, features or polish you never asked for, and saying "remove that, keep it simple" is part of the job. And know when to stop and restart rather than patching a run that went wrong.

People with management, editing or project experience are often better at this than they expect. The habits transfer directly.

## Example

A program lead asks AI to draft a volunteer handbook. Before any writing, she asks for an outline and cuts two sections. After the first chapter, she says "shorter sentences, and drop the legal language." By chapter three, the drafts need almost no edits.

## How it connects

- Extends **prompting** into an ongoing conversation about direction.
- Central to working with **agents** and to **vibe coding**.
- Cutting scope ties it to the **smallest useful version**; approvals tie it to **human in the loop**.

## Try it

On your next bigger AI task, ask for a plan first and don't let it start until you've changed at least one thing in the plan. Note what you changed and why.