Product1 publisher2 min readPublished
Pointing Claude Code at one directory made a developer's usage last four times longer
An XDA Developers writer says his Claude allowance stretched four times further once he stopped letting the agent orient itself inside a 500-file repository. The post does not include token counts, so read it as one practitioner's log.
The Product Desk · Product desk
What happened
- An XDA Developers writer says his Claude usage lasted four times longer once he stopped handing Claude Code his entire project at the start of a task.
- Opening a 500-file project does not push all of it into the context window; the window fills as the tool reads files, runs searches and commands, and takes in the output.
- Given a vague login-bug prompt, the post says the agent first orients itself, searching the repository for terms like login, auth, session or token before opening files.
- By the time it narrows the problem to two files, the post says it may have searched dozens of others and read code with nothing to do with the fix.
- The writer says that even when the change is a few lines in session.ts, you have paid for Claude working out half of your authentication system to get there.
Compiled by The Product DeskSomething wrong?How this is made
Why it matters
- cost The seat holder is funding rediscovery of things the team already knows, and the smallest fixes carry the largest overhead relative to the code that changes.
- decision A rate-limit complaint arrives as a purchasing question and can be answered as a practice one, so the scoped prompt is the cheaper thing to try before the upgraded plan.
- constraint A team that wants to defend this change internally has to instrument its own sessions first, because there is nothing in the account to check a 4x against.
- capability Clearing or compacting mid-task gives a developer a spending control inside the session, without waiting for anyone to approve a plan change.
The expensive part of a fresh session is orientation, and none of it comes back when a guess turns out to be wrong. Once the agent has searched for login, auth, session and token and opened whatever those searches surfaced, that material stays in the working memory unless the context is cleared or compacted, according to the post [7].
Two files out of 500 is 0.4 percent of the repository [14]. Dozens of others were searched to arrive at them [6]. Take the headline number at face value and the implied saving is steep: if the same allowance covers four times as much work, tokens per task drop to a quarter of what they were, a cut of about 75 percent [15]. The post carries no session logs, no stated method and no comparison between model tiers [13].
What an agentic session actually spends on, the writer says, is searches, the contents of the files those searches open, command output, and the conversation so far [4]. He puts the waste down to context, not to the difficulty of the change [12].
Before typing the task, it is worth pinning down which directory the problem is likely to be in and what reproduces it, and saying both if you know both. The writer does exactly that, mentioning that the failure only appears after a session expires and pointing at the authentication directory and the couple of files he suspects [10]. If you know neither, ask for a search first and read the result yourself instead of letting the agent start opening files, which is what he does when he cannot name the file [11].
This is for the person who has to explain a rate-limit wall to four developers on a Friday and whose only other lever is a bigger plan. The prompt in question is one sentence: "Start by looking in src/auth and only search elsewhere if you need to." The writer says even something that simple can dramatically narrow the amount of code Claude has to sift through [9].
What to watch
- Any first-party per-session breakdown of tokens spent on searches and file reads, which would let a team test the 4x claim against its own repository.
- Whether scope lines start appearing in CLAUDE.md and contributor docs as team policy rather than personal habit.
- Whether compaction defaults change so that irrelevant exploration drops out of the window without a manual clear.