PocketCoder-A1: Getting Claude to Work Three Shifts

Habr has published an article about PocketCoder-A1, a homemade "Auto-Coder" that the author built to make the Claude AI model work on a multi-shift schedule and get the most out of an existing LLM subscription.

The idea: let the AI work at night

The article starts from the author's observation that powerful AI models, in practice, do not so much replace human labor as increase its load. Instead of working less, specialists keep working just as much, only now alongside AI tools. The author's answer is to redistribute the load across time itself: if a person is busy during the day, the model can keep working at night, when the subscription would otherwise go unused.

This rests on something familiar to many who use subscription access to powerful LLMs: actual use of the limit over a 24-hour period is highly uneven. Peaks fall in working hours, while at night and on weekends access effectively sits idle. In economic terms, that is a missed opportunity. The user has already paid for the limit but does not use all of it, while any active codebase usually accumulates more tasks than needed that can be handed to a model without constant human supervision: refactoring, test generation, working through an issue backlog.

• Tool: PocketCoder-A1 ("Auto-Coder")
• Model: Claude
• Idea: "three shifts", meaning round-the-clock AI work instead of one-off use during working hours
• Motivation: get the most out of the LLM subscription, not out of individual requests
• Source: Habr

How does the economics of LLM subscriptions work?

Unlike pay-per-token API access, a subscription to an LLM creates an economic incentive to use the allotted limit as fully as possible: the user pays a fixed amount regardless of how much work the model actually does in that time. Tools such as Claude Code, which run on top of such a subscription, can set the model to work on tasks without constant human involvement. In theory, that means you can set up a task queue that the model works through in the background, including at night, when the person is not using the limit themselves.

This differs from classic API use billed by token count, where every extra request directly raises the bill and the user, conversely, has an incentive to minimize calls to the model. A subscription flips that logic. That is exactly why tools that automate background, round-the-clock model runs are especially attractive to people who already pay for a fixed-limit subscription rather than being billed by actual usage.

What does this mean for developers?

The author describes the tool through the metaphor of "three shifts": a mode in which the model works not occasionally, on request, but almost continuously, like an employee on a shift schedule. For a solo developer or small team, this is a way to raise effective productivity without raising subscription costs. If the limit is already paid for, idle hours become wasted compute that can be filled with background tasks such as refactoring, test generation and backlog processing.

Homemade "auto-coders" like this reflect a broader shift in how developers use commercial LLM subscriptions: not as a one-off chat assistant, but as a constantly running background process built into the development loop. How sustainable this approach is under the terms of use of specific subscriptions and their limits on automated requests is a question that, in homemade setups like this, usually stays at the author's own discretion and risk.

The author's own phrasing, "AI does not replace people, people just work more", captures an observation increasingly heard in discussions of AI's effect on development: the expectation that automation will free up specialists' time often ends up meaning they are expected to deliver more work in the same time, only now with the help of AI tools. In that sense, PocketCoder-A1 is less a technological solution than one developer's personal strategy for offsetting this growing workload, by handing at least part of the work to the model during hours when he physically cannot work himself.