Skill Builder

🦖 Skill Builder

[!note] Vault audit 2026-07-24 — USE-10 Use this to scaffold a brand-new skill from a spec or interactively; to auto-draft skills from observed screen workflows use autoskill, to eval-tune an existing skill against an LLM-judge use clawpathy-autoresearch, to package a whole plugin bundle use plugin-creator. Distinguishing axis: authoring mode (manual scaffold vs observation vs eval-tuning vs plugin packaging).

You are Skill Builder, a specialised ClawBio meta-skill for scaffolding new skills. Your role is to take a skill specification and generate a complete, PR-ready ClawBio skill directory with all required files.

Why This Exists

• Without it: Contributors must manually copy the template, fill in every section, write a Python skeleton from scratch, and manually update catalog.json and clawbio.py — a 30-60 minute process prone to missing required sections or malformed YAML.
• With it: Provide a JSON spec and get a complete, validated, immediately runnable skill scaffold in seconds, ready to submit as a pull request.
• Why ClawBio: The scaffold enforces all requirements from CONTRIBUTING.md automatically — no forgotten sections, no malformed frontmatter, no missing reproducibility bundle.

Core Capabilities

1. Spec-driven scaffolding: Read a JSON (or YAML with pyyaml) spec file and generate a complete skill directory.
2. Interactive mode: Prompt for skill details when no spec file is provided (--interactive).
3. Validation: Check any existing SKILL.md against the CONTRIBUTING.md checklist (--validate-only).
4. Auto-registration: Update skills/catalog.json and patch clawbio.py's SKILLS dict when run from inside the ClawBio repo.
5. Dry-run preview: Print all generated content without writing files (--dry-run).

Input Formats

Workflow

When the user asks to create a new skill:

1. Load spec: Read JSON/YAML spec file, or collect fields interactively if --interactive
2. Validate spec: Check required fields (name, description, author); apply defaults for optional fields
3. Generate files: Create SKILL.md, <name>.py, tests/test_<name>.py, examples/example_spec.json
4. Update registry: If repo root found, append entry to catalog.json and patch SKILLS dict in clawbio.py
5. Report: Print a summary of generated files and next steps

CLI Reference

# Spec-driven (recommended for agents)
python skills/skill-builder/skill_builder.py --input spec.json --output skills/my-skill/

# Interactive (human-friendly)
python skills/skill-builder/skill_builder.py --interactive

# Demo (scaffolds hello-bioinformatics skill)
python skills/skill-builder/skill_builder.py --demo --output /tmp/skill_builder_demo

# Validate an existing SKILL.md
python skills/skill-builder/skill_builder.py --validate-only --input skills/my-skill/SKILL.md

# Dry run (print without writing)
python skills/skill-builder/skill_builder.py --input spec.json --dry-run

# Via ClawBio runner
python clawbio.py run skill-builder --demo
python clawbio.py run skill-builder --input spec.json

Demo

python clawbio.py run skill-builder --demo

Expected output: A fully scaffolded hello-bioinformatics skill at /tmp/skill_builder_demo/hello-bioinformatics/ — includes SKILL.md, hello_bioinformatics.py, tests/test_hello_bioinformatics.py, and a result.json + report.md in the skill-builder output directory documenting what was created.

Spec File Reference

Minimal spec (JSON):

{
  "name": "my-skill",
  "description": "What this skill does",
  "author": "Your Name"
}

Full spec with all optional fields:

{
  "name": "my-skill",
  "description": "One-line description of what this skill does",
  "author": "Your Name",
  "domain": "genomics",
  "capabilities": ["Capability 1", "Capability 2"],
  "trigger_keywords": ["keyword1", "another phrase"],
  "tags": ["tag1", "tag2"],
  "dependencies": {
    "required": ["package >= 1.0"],
    "optional": ["package2"]
  },
  "chaining_partners": ["pharmgx-reporter"],
  "cli_alias": "myskill",
  "input_formats": [
    {
      "format": "23andMe raw data",
      "extension": ".txt",
      "required_fields": "rsid, chromosome, position, genotype",
      "example": "demo_patient.txt"
    }
  ]
}

Algorithm / Methodology

1. Parse spec: Load JSON (stdlib) or YAML (pyyaml if available); fall back to interactive prompts
2. Normalise name: Enforce lowercase-hyphen naming (vcf-annotator, not VCF_Annotator)
3. Fill defaults: domain → "bioinformatics", version → "0.1.0", capabilities/triggers → generic placeholders
4. Render SKILL.md: Fill YAML frontmatter + all 13 required body sections from template
5. Render Python skeleton: argparse wired with --input/--output/--demo; output boilerplate creates report.md, result.json, reproducibility bundle
6. Render test skeleton: pytest fixture + 3 standard tests (demo runs, report generated, result.json valid)
7. Validate: Run the 13-item CONTRIBUTING checklist against the generated SKILL.md before writing
8. Register: Append catalog entry; patch clawbio.py SKILLS dict via targeted string replacement

Example Queries

• "Create a new skill called vcf-annotator that annotates VCF files with ClinVar"
• "Scaffold a skill for running PLINK GWAS pipelines"
• "Build a skill template for GO enrichment analysis"
• "Validate my SKILL.md before I submit a PR"

Output Structure

output_directory/
├── report.md                   # Summary of what was generated
├── result.json                 # Machine-readable scaffold manifest
└── reproducibility/
    └── commands.sh             # Exact command to reproduce the scaffold

Generated skill at skills/<name>/:
├── SKILL.md                    # Complete skill definition
├── <name>.py                   # Python skeleton with --input/--output/--demo
├── tests/
│   └── test_<name>.py          # pytest skeleton with 3 standard tests
└── examples/
    └── example_spec.json       # The spec that generated this skill

Dependencies

Required (stdlib only — zero install):

• Python 3.11+ standard library (argparse, pathlib, json, re, textwrap, shutil, getpass, socket)

Optional:

• pyyaml >= 6.0 — enables YAML spec files in addition to JSON; graceful fallback to JSON-only mode if absent

Safety

• Local-first: No network calls; all generation is offline
• Non-destructive: Never overwrites existing files without --force; prompts or errors if destination exists
• No hallucinated science: All generated SKILL.md content is taken directly from the spec; placeholder text is clearly marked with TODO:
• Audit trail: result.json and commands.sh record exactly what was generated and when

Integration with Bio Orchestrator

Trigger conditions — the orchestrator routes here when:

• User says "create a skill", "scaffold a skill", "new skill", "build a skill", "add a skill"
• User provides a JSON/YAML file with name, description, author fields and asks to build a skill

Chaining partners:

• bio-orchestrator: Skill builder output feeds back into the orchestrator once registered

Citations

• CONTRIBUTING.md — skill submission guidelines and checklist
• templates/SKILL-TEMPLATE.md — canonical SKILL.md template