Anthropic skill-creator: a controlled CSV-skill trial

Tested and written by Fred at SkillsTested · Published 2026-10-11 · Updated 2026-10-11

We installed a pinned copy of Anthropic's skill-creator into a temporary Codex project, explicitly asked the agent to use it, and got an installable CSV-audit Skill. The generated helper gave the expected reports on our fixed samples and three extra edge cases. This is evidence of one directed workflow, not proof of automatic activation or better output than a no-skill run.

The job

Create a reusable Agent Skill to audit a local CSV with case_id, owner, and score columns. It must emit JSON with source row numbers and flag missing columns, repeated IDs, blank owners, and scores outside the integer range 0–10. We supplied two small, synthetic files and required validation plus a packaged .skill archive. There were no real customer records.

Download the passing CSV, the failing CSV, or the generated skill package. The package is a trial artifact for inspection, not a production endorsement.

What happened

1. We inspected the upstream folder at the fixed commit, including its license, scripts, network and permission surface. Its authoring helpers are local Python scripts; the review found no bundled credential requirement or network call for this task.
2. The standard skills add command copied skill-creator into the temporary project's .agents/skills/ directory. The trial prompt then explicitly told Codex to follow that installed workflow.
3. Codex wrote SKILL.md, a Python standard-library CSV helper, and two evaluation prompts. It ran the helper, the upstream quick validator, and the upstream packager.
4. The first validator attempt failed because the local system Python lacked PyYAML. With PyYAML 6.0.3 installed in an isolated virtual environment, validation and packaging passed. The generated CSV helper itself needs no third-party package.

Checks we independently repeated

• The passing file checked two rows, reported zero errors, and exited 0; a quoted comma in an owner name parsed correctly.
• The failing file reported three errors: repeated A-01 and score 11 on line 3, plus a blank owner on line 4. It exited 1.
• A missing required header stopped row checks; a multiline field kept the next record's physical line number; an unclosed quote produced a CSV parse error.
• The .skill archive passed ZIP integrity inspection and contains only the generated instructions and helper. Evaluation fixtures were kept outside the installable package.

Where this test stops

We tried a scoped Claude Code run once, but its API returned an organization-disabled error before the task began. Claude Code activation and output are therefore unverified. The Codex run was explicitly directed to use the skill, so it does not test whether a natural user request triggers it. We ran one agent attempt, no no-skill baseline, no independent evaluator, no description optimization, and no human review of output quality. We verified the CSV checks listed above, not every possible CSV dialect or production dataset.

For the general structure of a portable skill, read how to build a skill. Inspect the pinned upstream source before installing a newer version.