Feed a coding agent the smallest sufficient context, express requirements as machine-checkable contracts, and close the loop with a sub-second local verification command.
- A coding agent and one Python module to change
- A fast local test command you trust
Providing an entire repository dump to a coding agent increases noise, induces hallucinations, and dilutes attention. Effective AI coding relies on tight context budgets and bounded contracts.
The context window is an attention budget
Do not feed hundreds of unrelated files into the agent prompt. Provide only:
- The specific file to modify.
- The public interface signatures of direct callers and callees.
- The automated test file defining the desired behavior.
Express requirements as machine-checkable contracts
Natural language instructions like “make the API cleaner” lead to unpredictable rewrites. Instead, specify:
- Input types and output formats.
- Permitted dependencies and standard library constraints.
- Failure cases and explicit exception types.
- The exact verification command to run.
Fast feedback with local tooling
Ensure the agent can run local feedback loops using fast tools such as uv and pytest. A sub-second test loop allows the agent to iterate and fix errors autonomously before human review.
Review diffs for unintended side effects
Always inspect git diffs to ensure the agent did not delete unrelated comments, introduce unpinned dependencies, or modify shared global state.
Sources
The workflow relies on the uv documentation for fast environment and test commands, and on the pytest documentation for the verification loop; both are linked above where they are used.
Verification record
Editorial review of the budgeting model against the uv and pytest documentation. Verified 2026-09-02.
About the author
Organizational byline for FlyPython guides, verification records, and corrections. Editorial standards and contact details →