Skill Forge Judge

Skill Forge Judge

Evaluate any LLM-consumed prompt against quality standards, focused on knowledge delta, instruction clarity, and practical usability.

Core Philosophy

Good Prompt = Expert-only Knowledge − What Claude Already Knows

Restating defaults is token waste.

Three Types of Knowledge

Good prompt: >70% Expert, <20% Activation, <10% Redundant.

Evaluation Dimensions

Dimensions are grouped. Universal dimensions always apply. Type-specific modules apply based on what the prompt is. Multiple groups can apply to a single prompt.

Final grade = total score / total applicable points (as %)

Group U: Universal (80 pts) — always scored

MANDATORY — READ references/universal-dimensions.md

Group S: Skill Module (40 pts) — SKILL.md targets only

MANDATORY — READ references/skill-dimensions.md

Group C: CLAUDE.md / System Prompt Module (40 pts) — prompts that govern LLM behavior settings

MANDATORY — READ references/claude-md-dimensions.md

Group B: Bash/Shell Module (30 pts) — prompts that contain shell/CLI guidance

MANDATORY — READ references/bash-dimensions.md

Evaluation Protocol

Evaluator's Lens: Before reading the target, adopt this question: "Does Claude already know this?" — every section gets marked [E], [A], or [R] before scoring begins.

Step 0: Detect Prompt Type

Read the target and identify which groups apply:

[ ] Is it a SKILL.md file?              → Score U + S. Do NOT load claude-md-dimensions.md or bash-dimensions.md.
                                          MANDATORY: Load agentskills spec via Spec Cache (below) before scoring S1.
[ ] Is it a CLAUDE.md / system prompt?  → Score U + C. Do NOT load skill-dimensions.md.
[ ] Does it contain bash/shell rules?   → Also score B. Load bash-dimensions.md.
[ ] Is it something else?               → Score U only. Do NOT load any type-specific reference.

Multiple groups can apply (e.g. a CLAUDE.md with bash guidance → U + C + B).

Edge cases:

• SKILL.md that also contains bash guidance → U + S + B
• A referenced sub-file (e.g. references/bash.md, not a root prompt) → U only; note "sub-file, not root prompt" in report
• Ambiguous type (could be CLAUDE.md or skill) → score both C and S groups; note the ambiguity in the Summary

Spec Cache

The agentskills.io specification is cached daily so evaluations don't pay a network fetch on every run.

Cache dir:  ~/.claude/tmp/
Filename:   agentskills-spec-YYYY-MM-DD.md   (today's date, e.g. agentskills-spec-2026-04-14.md)

1. Glob ~/.claude/tmp/agentskills-spec-<today>.md
2. If found  → Read it — done.
3. If not    → Bash: mkdir -p ~/.claude/tmp
               WebFetch https://agentskills.io/specification
               If WebFetch fails → score S1 from training knowledge; note "spec unavailable" in S1 Notes cell — done.
               Write result to ~/.claude/tmp/agentskills-spec-<today>.md
               Read the file — done.

Only fetch when the target is a SKILL.md (S1 scoring). Skip entirely for CLAUDE.md / system prompt / other evaluations.

Step 1: First Pass — Knowledge Delta Scan

Read completely. Mark each section [E] Expert | [A] Activation | [R] Redundant. Calculate E:A:R ratio — target >70% Expert.

Step 2: Structure Analysis

Note prompt type(s), applicable groups, length, reference files, and loading/trigger mechanisms.

Step 3: Score Each Applicable Dimension

Before opening the rubric, ask: what does this section assume Claude doesn't already know?

Load the reference file for each applicable group. For each dimension: find specific evidence, assign score with one-line justification, note improvements if score < max.

If a reference file cannot be read, halt and report: [ERROR] Cannot score Group X — reference file not found: <path>. Do not proceed with that group.

Step 4: Calculate Score & Grade

Grade = (sum of scored dimensions) / (sum of applicable maxes) → apply grade scale.

Step 5: Generate Report

# Prompt Evaluation Report: [Name]

## Executive Summary
[1–2 sentences max. Overall verdict + one top strength or weakness. No scores, no tables, no lists — prose only.]

## Summary
- **Type**: [Skill / CLAUDE.md / System Prompt / Other] + applicable groups
- **Score**: X / Y (Z%)
- **Grade**: [A/B/C/D/F]
- **Knowledge Ratio**: Expert:Activation:Redundant = X:Y:Z
- **Verdict**: [One sentence assessment]

## Group Scores

| Group | Score | Max | % |
|-------|-------|-----|---|
| U: Universal | | 80 | |
| S: Skill | | 40 | | ← if applicable
| C: CLAUDE.md | | 40 | | ← if applicable
| B: Bash/Shell | | 30 | | ← if applicable
| **Total** | | | |

## Dimension Scores

Include only rows for groups detected in Step 0. Omit rows for groups that don't apply.

| ID | Dimension | Score | Max | Notes |
|----|-----------|-------|-----|-------|
| U1 | Knowledge/Instruction Delta | | 20 | One-line justification + primary gap if score < max |
| U2 | Mindset + Procedures | | 15 | One-line justification + primary gap if score < max |
| U3 | Anti-Pattern Quality | | 15 | One-line justification + primary gap if score < max |
| U4 | Freedom Calibration | | 15 | One-line justification + primary gap if score < max |
| U5 | Practical Usability | | 15 | One-line justification + primary gap if score < max |
| S1 | Specification Compliance | | 15 | One-line justification + primary gap if score < max |
| S2 | Progressive Disclosure | | 15 | One-line justification + primary gap if score < max |
| S3 | Pattern Recognition | | 10 | One-line justification + primary gap if score < max |
| C1 | Behavioral Clarity | | 15 | One-line justification + primary gap if score < max |
| C2 | Scope Definition | | 15 | One-line justification + primary gap if score < max |
| C3 | Structural Organization | | 10 | One-line justification + primary gap if score < max |
| B1 | Rule Specificity & WHY | | 10 | One-line justification + primary gap if score < max |
| B2 | Anti-Pattern Coverage | | 10 | One-line justification + primary gap if score < max |
| B3 | Scope & Exceptions | | 10 | One-line justification + primary gap if score < max |

## Critical Issues
[Must-fix problems that significantly impact effectiveness]

## Detailed Analysis
[For each dimension scoring below 80%, provide:
- What's missing or problematic
- Specific evidence with line numbers
- Concrete suggestions for improvement]

## Numbered Improvements
1. [Highest impact improvement with specific guidance]
2. [Second priority]
3. ...

(skill-forge-hitl-compatible — invoke /skill-forge-hitl to step through each item.)

Common Failure Patterns

MANDATORY — READ references/failure-patterns.md

Extending skill-forge-judge

To add a new evaluation group, MANDATORY — READ references/extending-groups.md before proposing any new group. Do NOT load this file during a normal evaluation run.

Self-Application

skill-forge-judge can and should evaluate itself. The criteria must be self-consistent — if skill-forge-judge can't score well against its own rubric, the rubric is wrong.

Applicable groups: skill-forge-judge is a SKILL.md with no bash guidance → U + S (120 pts max).

Expected score: ≥B (80%+, ≥96/120). A score below B indicates the rubric has drifted from its own standards and the lowest-scoring dimensions should be revisited. A score of A (90%+) on self-evaluation is a warning sign — it may mean the criteria were written to fit the evaluator rather than the other way around.

Known limitation: self-evaluation can be gamed by writing criteria the evaluator happens to satisfy. When evaluating skill-forge-judge itself, pay extra attention to whether dimensions capture genuine expert knowledge vs. formalizing what any thoughtful evaluator would do anyway.

NEVER Do When Evaluating

• NEVER give high scores just because it "looks professional" or is well-formatted — formatting is cheap; rewarding it masks content gaps and trains authors to polish instead of improve INSTEAD: Score content for expert knowledge density, not visual polish.
• NEVER ignore token waste — redundant content dilutes expert signal and teaches authors that padding is acceptable; deduct consistently to create the right incentive INSTEAD: Deduct from U1 consistently regardless of overall quality; note the specific redundant lines.
• NEVER let length impress you — a 500-line prompt with 80% activation content is worse than a 50-line one with pure expert knowledge; length is effort, not value INSTEAD: Measure the Expert:Activation:Redundant ratio; a short high-Expert prompt outscores a long Activation-heavy one.
• NEVER skip mentally testing decision trees — plausible-looking trees often have unreachable branches or missing cases that only surface when traced INSTEAD: Trace each branch to verify it terminates with a clear action; flag any branch that dead-ends or loops.
• NEVER forgive explaining basics with "but it provides helpful context" — it sets a precedent that activation content earns tokens; future versions accumulate more INSTEAD: Mark the section [R] and deduct from U1; note the specific lines in Detailed Analysis.
• NEVER overlook missing anti-patterns — absence of a NEVER list usually means the author hasn't hit the failure modes yet, not that they don't exist INSTEAD: Note "NEVER list absent" as a Critical Issue and deduct from U3.
• NEVER assume all procedures are valuable — generic procedures (open, edit, save) inflate U2 scores without adding knowledge delta INSTEAD: Ask "would Claude follow this procedure without being told?" — if yes, mark [A] or [R], not [E].
• NEVER undervalue the description field for Skills — it is the only thing the agent sees before deciding whether to load the skill; poor description = skill never activates regardless of content quality, making all other work invisible INSTEAD: Score S1 harshly for vague or keyword-poor descriptions; a skill that doesn't activate is a failed skill regardless of body quality.
• NEVER compare percentage scores across evaluations without checking which groups were scored — U+S (120 pts) and U+C+B (150 pts) are different denominators; the percentages are comparable, the raw scores are not INSTEAD: Always note the denominator in the report (e.g., "89% on U+S = 107/120") so cross-evaluation comparisons are valid.
• NEVER place Numbered Improvements before Detailed Analysis — the reader needs context before recommendations; without it, improvements land without evidence and get dismissed or misapplied INSTEAD: Always write Detailed Analysis first, then Numbered Improvements immediately after.