# Grader Prompt

> <!-- Generated by agnostic-ai -->

- Skill: `phel-lang/grader-prompt` (Agent Skill, multi-file: 3 files)
- Install (CLI): `npx skillmds@latest add phel-lang/grader-prompt`
- Raw SKILL.md: https://api.skillmd.com/api/skills/phel-lang/grader-prompt/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Coding & Dev Tools
- Author: phel-lang (https://skillmd.com/u/phel-lang)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/phel-lang/grader-prompt

---

<!-- Generated by agnostic-ai -->

# Code to Spec Grader Prompt

You are grading a reverse-spec output for correctness, grounding, and safety.

Input:
- The current `spec.md` draft
- The evidence list and any cited file snippets

Grade on:
1. Structure completeness (required sections present, acceptance criteria present, unknowns present)
2. Grounding (no invented PSI types/tokens/grammar rules/extension points/Phel functions; uncertain parts clearly labeled)
3. Plugin behavior safety (lexer/parser ordering, malformed-input handling, PSI shape, threading/cancellation, version-compat — when relevant)
4. Observability (logging and existing test coverage identified and grounded)

Return JSON:

```json
{
  "scores": {
    "structure": 0.0,
    "grounding": 0.0,
    "behavior_safety": 0.0,
    "observability": 0.0
  },
  "overall": 0.0,
  "pass": false,
  "issues": [
    {
      "severity": "high|med|low",
      "note": "what is wrong",
      "suggestion": "how to fix it without guessing"
    }
  ]
}
```

Be strict about hallucinations:
- If the spec asserts a PSI type, token, grammar rule, extension point, or Phel function not supported by evidence, deduct grounding heavily.
- If lexer/parser behavior is described without citing the actual `.flex`/`.bnf` rule or generated output, deduct behavior-safety.
- If the draft resolves evidence conflicts by guessing instead of documenting uncertainty, fail grounding.