Claude Code Overestimates Time: Two Methods to Correct It

Claude Code often provides time estimates aligned with human benchmarks, sometimes significantly exceeding the actual durations observed with the help of an LLM. Two approaches are proposed to make these estimates more relevant: leveraging past durations and breaking tasks down into finer components while explicitly estimating on the LLM side. These methods, although improvable, aim to enhance reliability in team coordination.

Breaking Down Tasks and Explicitly Estimating on the LLM Side

The first method involves asking the agent to break down the functionality into subtasks and estimate the duration of each, specifying that the estimate should reflect the time of an LLM rather than that of a human. This approach does not require extensive historical data: it relies on a sufficiently precise breakdown to facilitate estimation. Typical subtasks include research, implementation of different parts, and testing, but it is recommended to go even further in specificity. The finer the granularity, the more accessible the estimation becomes, both for a human and for a model. The agent can draw on an implicit understanding of its own efficiency, particularly for research, which is significantly faster with an LLM than with manual searching. Once each duration is estimated, they can simply be summed to obtain a total estimate.

Anchoring Figures in Delivery History

Another approach involves creating an internal database of actual durations and linking it to the agent. Tracking can be done by documenting each task in Linear tickets or in a markdown file on Notion, precisely noting the start and end times for each intervention by the agent. After several deliveries and corrections, it becomes possible to create a dedicated skill that Claude Code can consult when a request for estimation arises. This database can contain examples such as 12 hours for feature A, 3 hours for feature B, or 4 hours for a bug fix. When faced with a new request, the agent compares the current task to these precedents to adjust its estimate. This method improves the relevance of responses, although it does not guarantee complete accuracy, as it remains impossible to foresee all obstacles in advance.

Why Proposed Durations Remain Aligned with Human Estimates

The gap between Claude Code's estimates and reality primarily stems from the model's training data. Claude was trained on human-sourced data from the web, such as blog posts detailing implementation durations or statistics from repositories at a time when code was written without AI assistance, before 2022. The benchmarks thus integrated reflect human rhythms: a feature might take four weeks, a bug fix two days. The estimates generated today extend these references and do not correspond to the working pace of an LLM.

Usage Observations: Significant Discrepancies on Common Tasks

Feedback indicates responses of three to four weeks for a single engineer when an agent is asked about a chatbot feature. The author frequently observes this type of estimation, while the same task can be completed in a day with the right setup. He estimates that most features can be accomplished in twelve hours if the specification is clear and expresses frustration at the lack of estimates in minutes. More broadly, he notes that, without specific framing, most code AIs provide durations that exceed those observed in practice.

What This Changes for Organization and the Future

The reliability of estimates influences team coordination: product managers and colleagues need a clear horizon, especially since some tasks condition the work of others. The author anticipates a gradual improvement in estimates as more data on durations obtained with LLMs becomes available online, while expecting some delay before achieving better accuracy. This topic is deemed important in a context where these tools are changing practices and where the ability to align humans and models becomes an essential skill. An infographic is mentioned as a summary of key points.