SkillAnything

Automatically generate production-ready Skills for any target — software, API, CLI tool, library, workflow, or web service. SkillAnything runs a 7-phase pipeline that analyzes your target, designs the skill architecture, implements it, generates test cases, benchmarks performance, optimizes the description, and packages for multiple agent platforms.

Quick Start

Fully automated (one command):

Give SkillAnything a target and it handles everything:
- "Create a skill for the jq CLI tool"
- "Generate a skill for the Stripe API"
- "Turn this workflow into a multi-platform skill"

The pipeline runs all 7 phases automatically. Results land in

sa-workspace/

The 7-Phase Pipeline

Phase 1: Analyze    → Detect target type, extract capabilities     → analysis.json
Phase 2: Design     → Map capabilities to skill architecture       → architecture.json
Phase 3: Implement  → Generate SKILL.md + scripts + references     → complete skill directory
Phase 4: Test Plan  → Auto-generate eval cases + trigger queries   → evals.json
Phase 5: Evaluate   → Benchmark with/without skill, grade results  → benchmark.json
Phase 6: Optimize   → Improve description via train/test loop      → optimized SKILL.md
Phase 7: Package    → Multi-platform distribution packages         → dist/

See

METHODOLOGY.md

Usage Modes

Auto Mode (default)

Runs all 7 phases end-to-end. Provide the target and SkillAnything does the rest:

Target: "the httpie CLI tool"
→ Analyzes httpie --help output, designs command structure, generates skill,
  creates tests, benchmarks, optimizes, packages for 4 platforms

Interactive Mode

Set

auto_mode: false

config.yaml

• Phase 1 → "Here's what I found about the target. Look right?"
• Phase 2 → "Here's the proposed skill architecture. Any changes?"
• Phase 3 → "Draft skill ready for review."
• ...continues with user feedback at each step

Single Phase Mode

Run any phase independently:

python -m scripts.analyze_target --target "jq" --output analysis.json
python -m scripts.design_skill --analysis analysis.json --output architecture.json
python -m scripts.init_skill my-skill --template cli --output ./out
python -m scripts.generate_tests --analysis analysis.json --skill-path ./out/my-skill
python -m scripts.run_eval --eval-set evals.json --skill-path ./out/my-skill
python -m scripts.run_loop --eval-set trigger-evals.json --skill-path ./out/my-skill --model <model>
python -m scripts.package_multiplatform ./out/my-skill --platforms claude-code,openclaw,codex

Configuration

Edit

config.yaml

pipeline.auto_mode

true

target.type

auto

platforms.enabled

platforms.primary

eval.max_optimization_iterations

obfuscation.enabled

false

See

references/schemas.md

Platform Output

~/.claude/skills/<name>/

~/.openclaw/skills/<name>/

~/.codex/skills/<name>/

.skill

See

references/platform-formats.md

Evaluation and Benchmarking

SkillAnything uses the same eval system as the Anthropic skill-creator:

1. Test cases with assertions → graded by

   agents/grader.md

2. Benchmark comparing with-skill vs baseline →

   benchmark.json

3. Description optimization with train/test split → prevents overfitting
4. Interactive viewer via

   eval-viewer/generate_review.py

The eval loop is optional (

skip_eval: true

Scripts Reference

analyze_target.py

design_skill.py

init_skill.py

generate_tests.py

run_eval.py

aggregate_benchmark.py

generate_report.py

improve_description.py

run_loop.py

quick_validate.py

package_skill.py

package_multiplatform.py

obfuscate.py

Agents

Read these when spawning specialized subagents:

agents/analyzer.md

agents/designer.md

agents/implementer.md

agents/grader.md

agents/comparator.md

agents/optimizer.md

agents/packager.md

Target Types

SkillAnything auto-detects the target type and adapts its analysis:

Troubleshooting

• Phase 1 fails: Target not found or inaccessible → provide

  --target-type

  override
• Low eval scores: Description too vague → run Phase 6 optimization
• Platform packaging errors: Missing required fields → check

  references/platform-formats.md

• PyArmor not found: Install with

  pip install pyarmor

License

MIT License. See

NOTICE