create-slop-grader-rules

Files (1)

• SKILL.md

SKILL.md (first lines)

# Create slop-grader rules

Help the user create custom rule files (`.md`) for [`slop-grader`](https://github.com/lukstei/slop-grader/blob/main/README.md).

For complete syntax rules and validation constraints, refer to [docs/SYNTAX.md](https://github.com/lukstei/slop-grader/blob/main/docs/SYNTAX.md).

## Example Ruleset

```markdown
Custom documentation standards.

# Document Rules

## task_orientation
Is the guide organized around user tasks rather than internal system architecture?

### Criteria
- Architecture dump with no clear user workflow
- Mixed explanations requiring the reader to assemble the workflow
- Task-oriented with clear sequential steps
- Cookbook-grade with inputs, steps, and expected outputs

# Line Rules

## minimizing_language
Does the line use condescending shortcuts ("simply", "just", "obviously") to minimize a complex step?

### Criteria
- **true**: The line minimizes a step requiring configuration or technical decisions.
- **false**: The step is genuinely trivial (e.g. "click Save"), or no minimizing term appears.
```

## Core Principle: Keep Rules Granular

Design rules with single, atomic responsibilities. Avoid bundling multiple distinct checks into one composite rule (e.g. never combine typos, compound spacing, and grammar into a catch-all rule):

- **Debuggability:** When `slop-grader` flags a violation, the report displays the rule ID. Granular IDs (e.g. `compound_spacing`, `comma_placement`) pinpoint the exact issue instead of requiring manual deduction.
- **Evaluator accuracy:** Evaluators score focused, binary criteria far more reliably than composite prompts that test multiple disparate error types simultaneously.
- **Independent tuning:** Isolating rules lets you refine criteria and edge-case examples for one pattern without causing regressions in others.
- **Actionable fixes:** Downstream fix workflows (like the `slop-grading` skill) can propose exact replacements when the rule name is specific.

## Workflow

1. **Clarify requirements:** Ask what patterns, guidelines, or quality criteria the user wants to enforce. Determine the appropriate scopes:
   - `# Line Rules`: Evaluates line by line. Always boolean `noul` questions requiring `- **true**:` and `- **false**:` criteria.
   - `# Document Rules`: Evaluates the whole document. Multi-tier `score` questions (bullet list of at least 2 levels) or document-level boolean `noul` questions.
2. **Draft the ruleset:** Write the `.md` rule file adhering to [docs/SYNTAX.md](https://github.com/lukstei/slop-grader/blob/main/docs/SYNTAX.md).
   - Keep rules atomic (one pattern per rule ID). Split broad checks into separate rules.
   - Use descriptive snake_case for rule IDs (`empty_adverb`, `compound_spacing`).
   - Write unambiguous prompt instructions and explicit criteria bullets.
3. **Validate the rule:** Always run the validation command:
   ```sh
   npx @lukstei/slop-grader@latest --check -r <rule-file>
   ```
   If validation fails, fix the syntax errors until it passes.
4. **Instruct the user:** Present the final rule file and show the user how to run `slop-grader` against their documents:
   ```sh
   npx @lukstei/slop-grader@latest -r <path-to-rule-file> <path-to-target-document>
   ```
   **Important:** Never invoke the grading command yourself. The agent only validates rules via `--check`. The user runs the grading command directly.

From lukstei/slop-grader, MIT.

More from lukstei/slop-grader