Explicit Prohibitions

Instruction files must contain at least one constraint atom — a sentence that tells the agent what NOT to do. Without explicit prohibitions, the agent defaults to its training priors, which may include destructive actions like force-pushing, deleting files, or modifying sensitive configurations.

Antipatterns

• Only positive directives like "Use ruff for formatting" and "Run tests before committing" with no constraints — the check requires at least one sentence classified as a constraint (charge -1).
• Soft preferences instead of prohibitions like "Prefer not to modify the database" — hedged language does not register as a constraint atom.
• Prohibitions only in comments or code blocks like constraints inside <!-- --> or fenced blocks — the content query analyzes parsed instruction atoms, not raw text in non-prose regions.

Pass / Fail

Pass

Use `ruff` for formatting and linting.
*Do NOT run `black` or manual formatting.*
NEVER modify `.env` files directly.

Fail

Use `ruff` for formatting and linting.
Run `uv run pytest` before committing.
Follow the project conventions.

Limitations

Checks for at least one constraint atom (NEVER, MUST NOT, DO NOT). Does not evaluate whether the prohibitions cover the project's actual risk areas.