Agent Skill Creator

What It Does

Agent Skill Creator is a meta-skill: an agent skill whose job is to create other agent skills. You describe a workflow in plain English (or hand it a PDF, a URL, an existing script, or even a screenshot), and it produces a complete, validated, security-scanned skill directory with functional Python code, an eval spec, a cross-platform installer, and a README. The generated skill installs on 17 agent platforms including Claude Code, Cursor, GitHub Copilot, Gemini CLI, Windsurf, and more. The skill targets developers, data analysts, and teams who want to codify repetitive workflows into reusable agent instructions without writing spec files by hand.

The Good

The test suite is impressive. 153 unit tests covering validation, security scanning, eval rollouts, schema drift, dependency health, template structure, artifact detection, export parity, and v4 regression. All 153 pass under Python 3.12 in 4.19 seconds. CI runs the full matrix across Python 3.10 through 3.14.

Three bundled example skills all work end to end. The weekly-crm-report example takes a CSV with duplicates, deduplicates it, and produces correct regional totals (verified: 4 rows in, 3 deduped, West=100.50, East=350.25, grand=450.75). The pr-blocker-summarizer produces valid JSON summaries of PR states. The stock-analyzer generates RSI and MACD signals from mock data. All 27 rollout evals (9 per skill) pass with zero failures.

The quality gates are real, not decorative. validate.py checks frontmatter, naming, structure, and line counts. security_scan.py catches hardcoded tokens, eval/exec usage, and injection patterns. check_pipeline.py verifies scripts compile and declare dependencies. staleness_check.py tracks review dates, dependency health, and API schema drift. Every generated skill is supposed to pass all four before delivery.

Cross-platform support is thorough and honest. The skill documents 17 platforms across three tiers: native SKILL.md readers (Claude Code, Copilot, Codex CLI, Gemini CLI, Kiro, Cline, Roo Code, and others), auto-adapted platforms (Cursor gets .mdc, Windsurf gets .md rules, Trae gets .md with frontmatter, Junie gets guidelines.md), and manual platforms (Zed, Augment, Aider, Continue.dev). The install.sh dry-run correctly detected three platforms on this test machine.

Documentation is deep and well-structured. The repo ships 12,651 lines across SKILL.md, README.md, 15 Python scripts, and 14 reference documents. The README covers quickstart, comparison tables, examples, advanced install, platform tables, quality gates, staleness detection, team registry, troubleshooting, and project structure. Each reference file has a clear purpose listed in the SKILL.md reference table.

The Bad

The SKILL.md violates its own 500-line guideline. At 755 lines, the main SKILL.md exceeds the limit it prescribes for generated skills (the validate.py warns about this). The skill acknowledges the problem in its own validation output. For a meta-skill this comprehensive, the excess is understandable, but it is still a self-consistency issue.

Python 3.10 is the real minimum, not stated clearly upfront. The scripts use X | Y union type syntax (PEP 604) which fails on Python 3.9. macOS ships Python 3.9.6. The CI matrix starts at 3.10, which confirms the requirement, but the README quickstart does not mention it. The bootstrap script does not check the Python version either. On the test machine, check_pipeline.py crashed with AttributeError: module 'sys' has no attribute 'stdlib_module_names' under 3.9, and six test files failed to collect.

The security scan flags its own tooling. Running security_scan.py on the repo itself produces 10 findings: 4 high (eval/exec in security_scan.py itself) and 6 medium (false-positive secret patterns in documentation and test fixtures). This is a known pattern where security tools flag themselves, but it means the "block delivery if security fails" rule would block this very skill. The generated skills will not have this problem, but the meta-skill would fail its own gate.

No AGENTS.md companion file exists yet. The validate.py warns about this, and the skill instructions say every generated skill should produce one. The meta-skill itself does not ship one, which is another self-consistency gap.

Smoke Test Results

Tested on macOS 15 (Sequoia), aarch64, Python 3.9.6 (system) and Python 3.12 (homebrew).

Run A. Fresh clone, system Python 3.9.6

$ cd /tmp/agent-skill-creator
$ python3 scripts/validate.py .
✅ Status: VALID (6 warnings)

$ python3 scripts/security_scan.py .
❌ ISSUES FOUND (10): 4 high, 6 medium

$ python3 scripts/check_pipeline.py .
❌ AttributeError: module 'sys' has no attribute 'stdlib_module_names'

$ python3 -m pytest scripts/tests/ -q
❌ 6 collection errors (union type syntax unsupported)

$ cd references/examples/weekly-crm-report
$ python3 scripts/run_pipeline.py --input evals/golden/case-1/input.csv --output /tmp/summary.json
✅ Exit 0, correct JSON output

$ python3 scripts/run_evals.py --rollout
✅ 9 passed, 0 failed

Pass rate: 3 of 5. The pipeline and eval scripts work on 3.9. The validation tool works. The security scanner flags itself. check_pipeline and the test suite require 3.10+.

Run B. Same clone, Python 3.12 (uv run)

$ uv run --python 3.12 pytest scripts/tests/ -q
✅ 153 passed in 4.19s

$ python3.12 scripts/validate.py .
✅ Status: VALID (6 warnings)

$ cd references/examples/weekly-crm-report && python3.12 scripts/run_evals.py --rollout
✅ 9 passed

$ cd ../pr-blocker-summarizer && python3.12 scripts/run_evals.py --rollout
✅ 9 passed

$ cd ../stock-analyzer && python3.12 scripts/run_evals.py --rollout
✅ 9 passed, 1 llm-judge check (manual)

Pass rate: 5 of 5. Everything works under Python 3.12. All 27 automated eval rollouts pass. The llm-judge check for stock-analyzer correctly defers to manual evaluation rather than auto-grading.

Run C. Functional verification

$ cd references/examples/weekly-crm-report
$ python3 scripts/run_pipeline.py --input evals/golden/case-1/input.csv --output /tmp/summary.json
$ cat /tmp/summary.json
{
 "rows_in": 4,
 "rows_after_dedup": 3,
 "regions": {"West": 100.5, "East": 350.25},
 "grand_total": 450.75
}
✅ Correct: 4 rows in, 1 dup removed, West 100.50, East 200+150.25=350.25

$ bash install.sh --dry-run
✅ Detected: ~/.agents/skills/, ~/.claude/skills/, ~/.copilot/skills/

Functional pass rate: 3 of 3. Pipeline output matches expected values. Evals pass. Platform detection works correctly.

What the runs tell you

The skill is production-ready on Python 3.10 and above. All three example skills produce correct output. The eval framework catches real regressions. The only gap is the Python 3.9 compatibility failure, which affects the check_pipeline and test scripts but not the core skill content (SKILL.md is a doc-only instruction file). If you run Python 3.9 on macOS, upgrade to 3.10+ or install via homebrew.

Setup Walkthrough

1. Clone the repo: git clone https://github.com/FrancyJGLisboa/agent-skill-creator.git ~/.claude/skills/agent-skill-creator
2. Verify: cd ~/.claude/skills/agent-skill-creator && python3 scripts/validate.py . (should report VALID)
3. Or use the one-liner: curl -fsSL https://raw.githubusercontent.com/FrancyJGLisboa/agent-skill-creator/main/scripts/bootstrap.sh | sh
4. Open Claude Code and type /agent-skill-customer followed by your workflow description.

The install.sh script auto-detects your platforms. Use --dry-run to preview. The skill works immediately after cloning: no pip install, no API keys, no configuration.

Alternatives

1. Anthropic skill-creator - Interactive Q&A skill builder built into Claude Code. Simpler and faster for single-skill authoring, but Claude-focused, no cross-platform export, no eval specs, no security scanning. Good for quick one-offs.
2. Hand-writing SKILL.md - Direct authoring using the Agent Skills Open Standard. Maximum control, zero tooling dependency. The spec is not complex for simple skills. Best if you already understand the format.
3. K-Dense-AI/mimeo - Turns expert knowledge into SKILL.md or AGENTS.md files. Focuses on knowledge capture rather than workflow automation. Smaller scope, no pipeline or eval framework.