mk:rubric

Discovery, composition, and validation API for the rubric library at .claude/rubrics/. Consumed by the evaluator agent and mk:evaluate skill. Independently invokable via /mk:rubric SUBCOMMAND`` for manual inspection, validation, or composition outside the evaluator workflow.

What This Skill Does

• Lists all available rubrics and presets in the library
• Loads individual rubrics as prompt-ready fragments for injection into evaluator prompts
• Composes rubric presets (frontend-app, backend-api, cli-tool, fullstack-product) with their member rubrics and weights
• Validates rubric schema conformance and preset weight sums
• Enforces balanced PASS/FAIL few-shot anchor examples to prevent evaluator bias

When to Use

Activate when:

• User runs /mk:rubric [subcommand]
• An evaluator subagent needs to load rubrics for grading
• Sprint-contract negotiation references a rubric by path
• CI validates rubric schema conformance after edits

Core Capabilities

Subcommands

Subcommand	Purpose	Output
list	List all available rubrics + presets	Table: name, weight_default, applies_to
load NAME``	Load a single rubric and emit prompt-ready fragment	Markdown block ready to inject into evaluator prompt
compose PRESET``	Load a composition preset and return all member rubrics + weights	Composed prompt fragment with weight table
validate [path]	Validate one rubric (or all if no path) against schema.md	PASS / FAIL with diagnostics
validate --preset [path]	Validate composition preset (weights sum to 1.0 +-0.01)	PASS / FAIL

Presets

Preset	Loads	Use for
frontend-app	product-depth, functionality, design-quality, originality	SPAs, MPAs
backend-api	product-depth, functionality, code-quality	APIs, services
cli-tool	functionality, product-depth, code-quality, ux-usability	CLI tools
fullstack-product	All 7 rubrics (ux-usability weighted 3x)	End-to-end products

Frontend default is pruned to 4 rubrics per Phase 2 v2.0.0 -- product-depth, functionality, design-quality, originality. Do NOT auto-load code-quality, craft, or ux-usability for frontend targets -- they overlap existing kit layers.

Rubric Schema

All rubrics MUST conform to .claude/rubrics/schema.md. The validator enforces:

• Required frontmatter fields (name, version, weight_default, applies_to, hard_fail_threshold)
• Required sections in order: Intent, Criteria, Grading, Anti-patterns, Few-Shot Examples

• =1 PASS + >=1 FAIL anchor example, balanced (+-1)

• File <=200 lines

Composition presets MUST have all weights summing to 1.0 +-0.01.

Output Schema

load NAME`` output:

## Rubric: {name} (weight: {weight_default}, hard_fail: {threshold})

{Intent paragraph}

### Criteria
{bullets}

### Grading
{table}

### Anti-patterns
{bullets}

### Few-Shot Examples
{PASS + FAIL examples, balanced}

compose PRESET`` output:

## Composition: {preset-name}

| Rubric | Weight | Hard-Fail Threshold |
|---|---|---|
| ... | ... | ... |

(All member rubrics inlined below)

Calibration (from references/calibration-guide.md)

Few-shot anchor examples are the highest-leverage part of a rubric. They must follow three rules:

1. Balance PASS and FAIL counts -- With 2-3 total examples: tolerance +-1. With 4+ total examples: exact equality required. Imbalanced counts produce positive bias (40-60% inflation observed in harness research).
2. Randomize presentation order -- Alternate PASS/FAIL by example number (PASS, FAIL, PASS, FAIL...). Models attend more to recent context.
3. Pull from real prior reviews when possible -- Synthetic examples are a lossy proxy. Source from tasks/reviews/*-verdict.md, the Anthropic harness article appendix, or anonymized QA failures. Tag synthetic examples with <!-- synthetic -->.

Anti-patterns are FIXED -- they trigger FAIL regardless of the surrounding criteria. Don't add subjective ones.

Arguments

/mk:rubric list
/mk:rubric load <name>
/mk:rubric compose <preset>
/mk:rubric validate [path]
/mk:rubric validate --preset [path]

Workflow

/mk:rubric <subcommand>
    |
    |-- list: run load-rubric.sh --list, display table
    |-- load: run load-rubric.sh <name>, emit prompt fragment
    |-- compose: run load-rubric.sh --preset <name>, emit composed fragment
    |-- validate: run validate-rubric.sh [path], report PASS/FAIL
    |-- validate --preset: validate weight sum = 1.0 +- 0.01

Usage

# List all rubrics
rubric/scripts/load-rubric.sh --list

# Load one rubric
rubric/scripts/load-rubric.sh design-quality

# Compose a preset (returns all member rubrics + weight table)
rubric/scripts/load-rubric.sh --preset frontend-app

# Validate every rubric file in the library
rubric/scripts/validate-rubric.sh

# Validate a specific rubric
rubric/scripts/validate-rubric.sh path/to/rubric.md

# Validate a preset's weight sum
rubric/scripts/validate-rubric.sh --preset path/to/preset.md

Path convention: All commands assume cwd is $CLAUDE_PROJECT_DIR. Prefix paths with "$CLAUDE_PROJECT_DIR/" when invoking from subdirectories.

Common Use Cases

• Composing the frontend-app preset for mk:evaluate to grade a generator-built SPA
• Validating a newly authored rubric against schema before committing
• Listing all available rubrics to discover what grading dimensions exist
• Checking that preset weight sums remain at 1.0 after changing a rubric's weight_default
• Adding a new rubric to the library: drop a .md file, run validate-rubric.sh, register weight in presets

Example Prompt

/mk:rubric compose frontend-app I need to evaluate a generator-built SPA. Load the frontend-app preset so I can grade it against product-depth, functionality, design-quality, and originality with balanced PASS/FAIL anchors.

Pro Tips

• Don't load all rubrics -- the evaluator should load only the relevant preset, not the whole library. Context efficiency matters.
• Weight drift: if you change a rubric's weight_default, all presets that reference it must be re-checked for sum=1.0.
• Adding a new rubric checklist: frontmatter with all required fields, all 6 required sections, balanced anchors, alternating PASS/FAIL, concrete artifact descriptions, validate-rubric.sh passes, add to preset frontmatter and re-validate weight sum, add to RUBRICS_INDEX.md.
• Re-calibrate per model upgrade: when a new model tier rolls out, re-run anchors through the new model to verify it agrees with the stated verdicts. The Phase 8 benchmark suite (mk:benchmark) automates this.
• Hard-fail semantic: if any rubric hits its hard_fail_threshold, the overall verdict is FAIL regardless of weighted score. Soft averages do not save weak dimensions.