Skip to main content

Task Sizing for AI Agents

Overview​

Finding the right scope for AI-assisted tasks is crucial. Tasks that are too large overwhelm AI agents and produce poor results, while tasks that are too small don't leverage their capabilities effectively.

"You develop a feel for what is the right size and scope of a task that is good to delegate to Claude. If you gave them all of Arc like implement all the features of Arc, it would not do a good job. If you said change the border radius of this button... probably too small and not worth the squeeze."

The Goldilocks Zone​

Too Large ❌​

  • "Implement the entire user authentication system"
  • "Build a complete dashboard with all features"
  • "Refactor the entire codebase to use TypeScript"

Why it fails: Overwhelming context, too many architectural decisions, high likelihood of inconsistencies.

Too Small ❌​

  • "Change the border radius from 4px to 6px"
  • "Fix this typo in a comment"
  • "Update a single CSS property"

Why it's inefficient: Overhead exceeds value, could be done faster manually.

Just Right ✅​

  • "Implement add participant UI component with state management"
  • "Create user profile editing form with validation"
  • "Add search functionality to the patient list view"

Why it works: Clear scope, requires both thinking and implementation, completable in a focused session.

Developing Intuition​

It Takes Practice​

"There's somewhere in between that I think takes some feel for getting used to. I think it also differs depending on front end and back end... you got to just try it for yourself and you develop a feel for what the right sort of scope is."

Building your intuition:

  • Start conservative with smaller tasks
  • Learn from experience - notice when tasks are too large (lots of back-and-forth) or too small (overhead exceeds value)
  • Frontend vs backend tasks may have different optimal sizing
  • Practice regularly - like any skill, it improves with use

Breaking Down Large Tasks​

Instead of "implement the entire add participant feature," break it down:

Stage 1: "Add an add participant UI component and the state management to flip between these two states"

  • Specific scope: UI component with state switching
  • Clear boundaries: Don't populate with real data yet
  • Manageable complexity: Focus on UI and state only

Why this works: Focused scope prevents AI from getting ahead of itself, clear boundaries, manageable chunks.

Context and Scope Management​

Managing AI ambition: "Sometimes Claude can be very ambitious and get a little bit ahead of itself. So if you're trying to intentionally reduce the scope, you might tell it don't populate this drop down with real data right now because that's not the scope of this change."

Effective scope communication:

  • Be explicit about boundaries
  • Prevent scope creep by clearly stating what NOT to include
  • Keep AI focused on the current step
  • Save improvements for later

Frontend vs Backend​

Frontend tasks can often be larger because:

  • Visual feedback makes progress easier to assess
  • Component boundaries provide natural scope limits
  • UI feedback helps validate progress incrementally

Backend tasks should generally be smaller because:

  • Business logic complexity can be hidden
  • Data integrity issues have higher consequences
  • Integration points are more complex

Signs of Good Task Sizing​

You've sized it right when:

  • Completable in a focused session (1-3 hours)
  • Clear success criteria that can be validated
  • Independent testability
  • Reasonable review scope
  • Natural stopping points for commits

You need to adjust when:

  • Too many back-and-forth iterations with AI (probably too large)
  • Trivial changes that take longer to explain than implement (probably too small)
  • Unclear when it's "done" (scope not well defined)