Estimates get weird with Claude Code. Tasks that used to take three hours now take one. But some tasks that seemed like they'd be fast are actually slower, because the bottleneck moved.
Here's what changed and how I account for it.
What got faster
Implementation code — the actual writing — is roughly 3-5x faster with Claude. Boilerplate, tests, documentation, refactoring: all dramatically faster.
If your old estimates were based mostly on typing time, cut them significantly.
What didn't get faster
Requirements clarity. The time you spend understanding what needs to be built doesn't shrink with Claude. If anything it matters more, because Claude will happily build the wrong thing faster than you could build it yourself.
Code review. You still have to read and evaluate everything Claude writes. Budget 20-30% of implementation time for review.
Integration and testing. Getting Claude's code to fit with the existing system often takes work. Claude doesn't know your conventions as well as you do; it fills in gaps with reasonable defaults that may not match.
Debugging Claude's mistakes. These are rare but time-expensive when they happen. Budget a small buffer.
Using Claude to estimate
I now use Claude to help scope tasks before estimating them:
I need to build [feature]. Before I estimate it:
1. What are all the subtasks involved?
2. What are the hidden dependencies?
3. What assumptions am I probably making that could be wrong?
4. What's the riskiest part — the part most likely to take longer than expected?This surfaces things I would have missed in a manual estimate. Claude is good at enumerating subtasks because it's seen many similar implementations.
The estimate prompt
After scoping:
Given this task breakdown, estimate each component:
- Time for me to clarify requirements and write the prompt
- Time for you to implement
- Time for me to review and iterate
- Time for integration and testing
Give a range, not a single number. What makes it shorter? What makes it longer?Ranges beat point estimates. The low end is if everything goes right. The high end accounts for the riskiest assumption being wrong.
The new bottleneck
With Claude Code, the bottleneck in most tasks is no longer implementation — it's specification. The time you spend getting requirements right, writing clear prompts, and reviewing output is now the majority of the time budget.
This means the developers who estimate best with Claude are the ones who budget generously for thinking, scoping, and review — not the ones who just divide their old estimates by the implementation speedup.
Task scoping prompts are in the Agent Prompt Playbook — including templates for breaking down tasks before estimating. $29.