docs: refine agent quality guidance

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
This commit is contained in:
David Kaya
2026-03-24 21:17:40 +01:00
co-authored by Copilot
parent f1a96773c8
commit 76846843ab
+3 -2
View File
@@ -34,6 +34,9 @@ These instructions apply to any automated or semi-automated agent working in thi
- Avoid duplication when a shared abstraction improves clarity and maintainability.
- Prefer composition and small reusable units over sprawling functions, deep inheritance, or tightly coupled code.
- Make invalid states difficult to represent through types, structure, or APIs whenever the technology supports it.
- When refactoring, actively remove code smells instead of relocating them. Large multi-responsibility files, long methods, scattered mutable state, duplicated logic, and protocol or reflection glue mixed into orchestration code should usually be split into focused collaborators.
- Keep coordinator and orchestration types thin. Push policy decisions, parsing or translation, projection or mapping, and mutable execution state into dedicated helpers or services with clear ownership.
- For deep refactors, add characterization tests around the current behavior before extracting logic, then keep the tests aligned with the new collaborators so structure improves without changing behavior accidentally.
- Add comments only when they explain intent, reasoning, or non-obvious behavior that the code itself cannot communicate clearly.
## 4. Core delivery standards
@@ -52,7 +55,5 @@ These instructions apply to any automated or semi-automated agent working in thi
- Use Bun for dependency management and script execution.
- Prefer repository-local tooling over global machine state whenever possible.
- Keep changes focused and reviewable. Avoid mixing unrelated concerns into a single change.
- When the user asks only for a plan, stay in planning flow: analyze the codebase, clarify scope as needed, and write or update the plan, but do not begin implementation until the user explicitly asks you to start the work.
- Do not trigger workflow or mode transitions that implicitly approve or begin implementation unless the user has explicitly requested that transition.
- Always commit completed repository changes before handing work off. If unrelated pre-existing changes are present in the worktree, stop and ask the user how to proceed before creating the commit.
- Do not mark work as done until both the implementation and its verification are complete.