auto-skill-forge（自动技能锻造）

Skill Forge

Generate Skills from requirements AND generate task_suite.yaml for evaluation.

The primary value of this skill is task_suite generation -- turning a SKILL.md into a structured test harness that improvement-evaluator can run. Secondary value is generating SKILL.md from a structured skill_spec.yaml.

Key differentiator: Skill Forge does not merely scaffold a skeleton; it performs static analysis of the SKILL.md to extract testable claims (from five distinct sources) and assigns the appropriate judge type per task, producing a suite that is immediately runnable by improvement-evaluator.

When to Use

1. Add tests to an existing Skill -- You have a SKILL.md but no task_suite.yaml.

Run --from-skill to analyze the SKILL.md and generate a test suite automatically. The generator extracts scenarios from five sections of the document and assigns the right judge per task.

1. Create a new Skill from scratch -- You have a skill_spec.yaml describing

what the skill should do. Run --from-spec to generate both a complete SKILL.md (with frontmatter, sections, examples) and a matching task_suite.yaml.

1. Full lifecycle -- Generate a skill, test it with --evaluate, and optionally

improve it with --auto-improve in one pass. Combines skill-forge, improvement-evaluator, and improvement-orchestrator into a single command.

1. Validate coverage of an existing suite -- Compare the generated task_suite

against a hand-written one to find untested scenarios.

When NOT to Use

• You just want to evaluate an existing skill with an existing task_suite.yaml

→ use improvement-evaluator

• You just want to improve a skill that already has tests

→ use improvement-orchestrator

• You want to manually write a SKILL.md with templates

→ use skill-creator

• You want to score improvement candidates (not generate tests)

→ use improvement-discriminator

• You want to merge overlapping skills into one consolidated skill

→ use skill-distill

Modes

Mode A: --from-skill (Generate task suite for existing SKILL.md)

# Generate test suite for an existing skill
python3 scripts/forge.py --from-skill /path/to/skill-dir --output /path/to/output

# Generate and immediately evaluate
python3 scripts/forge.py --from-skill /path/to/skill-dir --output /path/to/output --evaluate

Reads the SKILL.md, extracts scenarios from five sources in priority order:

1. Frontmatter (description, triggers) → 1 core-capability test
2. "When to Use" bullets → up to 3 positive use-case tests
3. <example> tags → up to 2 keyword-match tests (uses ContainsJudge)
4. <anti-example> tags → up to 2 negative tests (should avoid bad patterns)
5. Output format/CLI sections → 1 format compliance test

Produces task_suite.yaml in the output directory. Example generated output:

skill_id: release-notes-generator
version: "1.0"
generated_by: skill-forge
tasks:
  - id: release-notes-generator-core-capability
    description: "Test core capability described in skill description"
    prompt: "You are an AI assistant with this skill loaded..."
    judge:
      type: llm-rubric
      rubric: "The output should demonstrate the capability..."
      pass_threshold: 0.7
    timeout_seconds: 120
    source: frontmatter.description
  - id: release-notes-generator-use-case-01
    description: "Use case: Generate notes from git log between tags"
    prompt: "Scenario: Generate notes from git log..."
    judge:
      type: llm-rubric
      rubric: "The output should address this use case..."
      pass_threshold: 0.6
    timeout_seconds: 120
    source: when_to_use

Mode B: --from-spec (Generate skill + task suite from spec)

# Generate complete skill from a spec file
python3 scripts/forge.py --from-spec spec.yaml --output /path/to/output

# Generate, evaluate, and auto-improve if below SOLID grade
python3 scripts/forge.py --from-spec spec.yaml --output /path/to/output --auto-improve

Reads a skill_spec.yaml (see references/spec-format.md) and generates:

1. A complete SKILL.md with frontmatter, When to Use / When NOT to Use, examples, output format
2. A task_suite.yaml derived from the generated SKILL.md (same five-source extraction)

The spec format requires only name and purpose; optional fields (inputs, outputs, quality_criteria, domain_knowledge, reference_skills) enrich the generated SKILL.md. Example minimal spec:

name: release-notes-generator
purpose: Generate structured release notes from git commit history

inputs:
  - name: commits
    type: git-log
    description: "Git commit log between two tags"

outputs:
  - name: release-notes
    format: markdown
    description: "Structured release notes with sections"

quality_criteria:
  - name: completeness
    description: "All commits accounted for in the notes"
    weight: 0.3

Common Flags

Output Artifacts

All generated YAML files use allow_unicode: True and default_flow_style: False for human readability. Files are written atomically (write-then-rename) to prevent corruption on crash.

Task Generation Strategy

The generator extracts test scenarios from 5 sources in the SKILL.md:

1. Frontmatter description → 1 core-capability test
2. "When to Use" section → up to 3 positive use-case tests
3. <example> tags → up to 2 keyword-match tests
4. <anti-example> tags → up to 2 anti-pattern avoidance tests
5. Output format section → 1 format compliance test

Tasks are deduplicated and capped at 10 per suite.

Judge Selection

• Scenario with specific keywords/outputs → ContainsJudge
• Scenario requiring quality/style assessment → LLMRubricJudge
• Scenario with structured output → PytestJudge (if test script can be generated)

Harness Pattern Tasks (for scripted skills)

When the target skill has a scripts/ directory, forge auto-generates additional test tasks checking execution-harness pattern adoption:

• Timeout handling: Does the skill handle subprocess.TimeoutExpired?
• Atomic writes: Does it use write_json/write_text from lib/common (atomic write-then-rename)?
• Backup/rollback: Does it create backups before file modifications?
• Error escalation: Does it have graduated error handling (not just crash-on-first-failure)?
• State persistence: Does it write recoverable state for crash recovery?

These tasks use ContainsJudge to grep the skill's Python source code. They only apply to orchestration/tool-type skills — pure-text knowledge skills skip this category.

Why Null-Skill Calibration

A generated task suite is only useful if it measures the _skill's_ contribution, not the base LLM's general ability. Without calibration, a naive suite can report 80%+ pass rates even when the SKILL.md adds zero value -- because the LLM already knows how to answer those questions.

Null-skill calibration addresses this by running every candidate task against a "null skill" (empty context, no SKILL.md loaded). Any task the null skill passes trivially is filtered out before the final suite is emitted. This ensures that every surviving task genuinely requires the knowledge or structure encoded in the SKILL.md.

Tradeoff: Null-skill calibration adds one extra LLM call per candidate task (or a heuristic keyword check in --mock mode). For a typical 10-task candidate set, this means ~10 additional calls during suite generation. The cost is justified because an uncalibrated suite gives false confidence: a skill that scores 9/10 on easy tasks looks "SOLID" but may add no value over a bare model. Calibrated suites reliably distinguish genuine skill contributions from baseline LLM capability.

When --mock is used, calibration falls back to a heuristic: tasks whose prompt contains only generic verbs ("explain", "describe", "list") without skill-specific terminology are filtered. This is less precise than LLM-based calibration but costs zero API calls.

The calibration step runs after deduplication and before the final cap of 10 tasks per suite.

Related Skills