What to delegate to AI coding assistants and what to keep for yourself. Distilled from Anthropic's internal research: 132 surveys, 53 interviews, 200k transcripts.
Anthropic studied how their own engineers use Claude. The findings reveal clear patterns about what to delegate, what to keep, and how trust evolves over time.
The Numbers
60%
of work uses Claude
50%
productivity boost
27%
new work that wouldn't happen otherwise
2x
YoY increase in both metrics
The 6 Delegation Criteria
Engineers delegate tasks that match these criteria:
1
Easily Verifiable
You can quickly check if the output is correct.
"It's amazing for everything where validation effort isn't large compared to creation effort."
2
Low Stakes
Code quality isn't critical. Throwaway or research code.
"If it's throwaway debug or research code, it goes straight to Claude."
3
Well-Defined
Self-contained tasks with clear boundaries.
"If a subcomponent is sufficiently decoupled from the rest, I'll get Claude to take a stab."
4
Boring / Repetitive
Tasks you're resisting or don't enjoy.
"The more excited I am to do the task, the more likely I am to not use Claude."
5
Outside Your Expertise
Low context for you, but low overall complexity.
"I don't know Git or Linux very well... Claude covers for my lack of experience."
6
Faster to Prompt
Tasks where prompting beats doing it yourself.
"If I anticipate it'll take me less than 10 minutes... I'm probably not going to bother using Claude."
What to Keep
1
High-Level Design
Architecture decisions, strategic thinking, system design.
2
"Taste" Decisions
Choices requiring organizational context or judgment.
3
High-Context Tasks
Work where you have deep knowledge that's hard to transfer.
4
Critical Code Quality
Production code where standards matter most.
The boundary is a moving target. Engineers regularly renegotiate what they delegate as models improve. Design/planning usage grew from 1% to 10% in six months.
The Trust Progression
How engineers build trust over time
Basic questions
→
Unfamiliar domains
→
Familiar tasks
→
Trust the suggestion
One engineer compared it to Google Maps: "At first I used it only for routes I didn't know. Today I use it all the time, even for my daily commute. If it says to take a different way, I just trust it."
The Supervision Paradox
The skills you need to supervise AI may atrophy from using AI
Effectively using Claude requires supervision. Supervising Claude requires the very coding skills that may atrophy from AI overuse.
"When producing output is so easy, it gets harder to actually take the time to learn something."
"If I were earlier in my career, it would take deliberate effort to grow my own abilities rather than blindly accepting model output."
Some engineers combat this by deliberately practicing without AI: "Every once in a while, even if I know Claude can nail a problem, I will not ask it to. It helps me keep myself sharp."
The Full-Stack Effect
AI enables everyone to work outside their core expertise. Different teams use Claude to augment their skills:
Security
Analyzing unfamiliar codebases
Alignment & Safety
Building front-end data visualizations
Backend Engineers
Creating complex UIs
Non-Technical Staff
Debugging and data science
The "Papercuts" Pattern
8.6% of Claude Code tasks are small quality-of-life fixes that would typically be deprioritized: refactoring for maintainability, building small tools, fixing minor issues. These add up to larger productivity gains over time.