kernel-design-agents CLAUDE.md

Agent Instructions

This repository is the generic Kernel Design Agents workflow reference. It should stay small and task-agnostic.

Repository Rules

• Use English for repository-facing files, comments, documentation, prompts, and commit messages.
• Keep task-specific prompts, datasets, validators, generated implementations, benchmark logs, and candidate artifacts out of this repository.
• Treat benchmark competitions, including MLSys-style work, as downstream applications of KDA rather than as the scope of this repository.
• Put generated outputs in runs/, outputs/, or profile/; these paths are ignored by git.
• Prefer documenting reusable workflow mechanics over documenting one task's private harness or acceptance thresholds.

Expected Agent Workflow

For a new task:

1. Create or enter a separate implementation workspace.
2. Define the task objective, constraints, validation command, and promotion criteria.
3. Use prompts/basic-flow.md as the starter prompt.
4. Read local task code and documentation before proposing implementation changes.
5. Write the initial plan draft to docs/draft.md inside the task workspace.
6. Convert the draft into an executable plan.
7. Implement in small iterations, validating each meaningful candidate.
8. Record candidate relationships, evaluation results, and profiling evidence when applicable.
9. Keep this repository focused on the reusable flow.

Optional Skills

Use external skills only when they are relevant to the active task:

• humanize for plan generation and implementation loops.
• A domain knowledge skill for background research.
• A profiling or report-analysis skill for performance evidence.