$0.50 task budget: Claude Code, fewer file reads

01 / Start with a $0.50 spending target, not a repository tour

If you have $0.50 available for a small coding task, do not spend the first conversation asking Claude Code to understand your entire project. Give it one change, a likely location, and a reading boundary. The $0.50 here is a personal spending target, not a promised task price or an automatic billing limit.

File contents enter the conversation as tokens. Reading unnecessary code can increase input usage, and that context may affect later requests too. Caching can change the cost, but fewer files do not translate into a fixed percentage saving. Measure billed usage rather than guessing from file counts.

02 / Find one likely folder before asking for edits

Spend two minutes locating the problem yourself. Use your editor's filename search or search for a visible error message. You do not need to understand the implementation; you just need a better starting point than the repository root.

• ■ For a broken login form, start with the login component or authentication folder.
• ■ For a failing test, provide the test path and the failure output.
• ■ For a configuration problem, name the relevant configuration file and the command that failed.

Instead of saying, “Explore this repository and fix login,” say, “The login button stays disabled after a valid email is entered. Start in src/auth/LoginForm.tsx. Do not inspect unrelated features.” Replace that example path with a real path from your project.

03 / Set a three-file checkpoint in the first prompt

Give Claude Code a small initial reading allowance. Three files is a useful beginner checkpoint, not a universal optimum. One file might contain the bug; another might define a helper; a third might contain its tests.

Try this prompt: “Fix the disabled login button. First read src/auth/LoginForm.tsx and its direct test file, if present. Read no more than three files before reporting what you found. Before opening additional files, list their paths and explain why each is needed. Avoid full-repository searches unless the local evidence is insufficient.”

This is an instruction, not a security boundary or enforced spending cap. Watch the tool activity. If the investigation expands, interrupt and ask for a narrower plan. Permit another file when there is a clear reason; refusing necessary context can create a wrong fix and a more expensive second attempt.

04 / Put a short reading policy in CLAUDE.md

For repeated work, add a compact CLAUDE.md file at the project root. Keep it useful rather than turning it into another document the assistant must digest. Include the main source directory, test command, generated folders, and the preferred search scope.

For example: “Application code is in src/auth. Authentication tests are in tests/auth. Start with task-relevant paths. Search filenames before reading contents. Exclude generated output from exploratory searches. Ask before broadening beyond the named feature.”

Do not paste secrets into this file. Also, do not treat ignore rules as a guaranteed access barrier: search exclusions and tool permissions serve different purposes. Use permissions for sensitive access restrictions and reading instructions for economical investigation.

05 / Trade broad file dumps for targeted evidence

Ask for filename matches before content matches, and matching lines before whole large files. A bounded search inside src/auth is usually a better first move than searching every directory for “user.”

For illustration only, eight files averaging 1,500 tokens contain about 12,000 tokens. Two relevant excerpts averaging 800 tokens contain about 1,600. Those are assumed sizes, not a benchmark or a billing prediction. The excerpts could also omit something important, so expand deliberately when needed.

06 / Check the receipt before starting the next task

After the fix and its targeted test, inspect your provider's usage record. Track the task, billed amount, and whether the assistant wandered. Start a fresh session for unrelated work instead of carrying an old investigation forward.

CheapRelay fits this approach with pay-as-you-go billing, no monthly fee, and no subscription. Billing includes real tokens, including cache reads and writes, so cached context is not automatically free. Sign up for CheapRelay if you want to test this workflow on a small budget, then compare actual receipts before increasing your spending.