Independent reviews. IonBin has no partnerships with the tools on this page and was not paid for these listings. Our evaluations are our own, based on public documentation and our research. Product names are trademarks of their owners and are used for identification only. How disclosure works.

Runtime probedSkill authoringExplored

Anthropic / skills/skill-creator/SKILL.md

Skill Creator

Drafts and improves skills through example tasks, baseline comparisons, human feedback and trigger evaluations.

Read upstream skill GitHub repository Add to comparison

Runtime probed · reviewed · A helper or bounded runtime behavior was probed. This does not establish model task effectiveness or comprehensive security.

01 / Task fit

Where it may help.

These are intended uses indicated by the upstream material. They are not measured outcomes.

  • Turning a repeatable workflow into a skill
  • Comparing a skill revision with its baseline

02 / Setup & access

What it asks of your environment.

Requirements

  • Drafting needs an agent and writable workspace; automated trigger evaluation additionally uses Python 3 and the claude CLI.
  • The description improver uses the session's Claude Code authentication; its source explicitly says no separate ANTHROPIC_API_KEY is needed.
  • The live viewer uses Python's standard library, a browser and a local TCP port; its port cleanup invokes lsof when available. Static HTML output is supported.
  • The trigger runner uses select() on subprocess pipes; native Windows operation was not verified. Shell examples include Unix commands.

Permissions & data access

  • Create skill files, evaluation outputs and temporary .claude/commands entries.
  • Run Claude subprocesses that inherit the surrounding environment and can use the configured Claude service.
  • Serve evaluation artifacts locally; live-viewer startup can terminate processes on its chosen port.
Check the host, too.

Instructions alone do not enforce a sandbox. The agent host, connected tools and approvals determine what the skill can actually read, change or send.

03 / Things to check

Security concerns & practical limits.

These concerns follow from the reviewed sources. They describe possible exposure or required review; they are not confirmed vulnerabilities.

Higher attention

Evaluation text can escape the viewer script context

Observed in the bounded evaluation: At revision 34040c9c568585f6929bedeaad110ad08f079624, generate_html embeds json.dumps output directly in an inline script. Synthetic prompt text with a closing script tag executed a harmless in-memory marker in fresh offline Chrome. The benign control did not execute the marker. This is confirmed only for the pinned files and fixture; current upstream HEAD was not tested. Two failed expectations describe this one issue.

Higher attention

Viewer can terminate another process

generate_review.py calls lsof and sends SIGTERM to processes listening on its selected port before starting the viewer (default 3117). IonBin assessment: use static output or inspect the port before live mode.

Review before use

Evaluation data and consumption

The improver sends skill content and evaluation examples to claude -p; repeated runs use the configured Claude account. The viewer embeds outputs. IonBin assessment: use approved sample data and review cost and artifact exposure.

Limits and nuances

  • Selected runner, improver and viewer code paths reviewed; packaging, all graders and the full optimization loop were not audited.
  • A good score on a small task set does not establish general reliability or resistance to hostile inputs.

04 / Review record

What we checked. What’s still open.

Runtime probed

Recorded review scope

  • SKILL.md drafting, baseline-evaluation and description-optimization sections
  • scripts/run_eval.py subprocess launch, environment and pipe handling
  • scripts/improve_description.py Claude subprocess and prompt construction
  • eval-viewer/generate_review.py artifact embedding, port cleanup and localhost server paths
  • LICENSE.txt
  • Bounded evaluation; exact operations, counts, environment and limitations in the linked public report
Not verified

Not tested or established

  • No effectiveness benchmark or security certification
  • Transitive dependencies, current advisories and host-agent permission controls not audited
  • No generated-skill quality or trigger-rate measurements
  • No Claude subprocess, live model, API credential, optimization loop or generated-skill effectiveness evaluation
  • Port-cleanup lsof/os.kill effects were mocked; no real process was terminated
Review date
Source revision
34040c9c568585f6929bedeaad110ad08f079624
Our relationship
Explored. Investigated through the recorded review. No use or replication is claimed.

Upstream content can change after review. A repository revision identifies source scope; it does not prove that every script, dependency or runtime behavior was tested. How freshness is handled.

Recorded evaluation / Helper probe

Results with their limits attached.

Completed within scope

Viewer generation, mocked port cleanup and offline browser marker: 5 of 7 expectations passed. Two checks confirm one script-context injection issue at the pinned revision.

Passed
5
Failed
2
Total recorded
7

These counts refer to the cases in this report. They are not a benchmark against other skills. A helper probe does not measure model task effectiveness or establish comprehensive security.

Read the evidence report

Download the authored reproduction harness (JSON bundle) ↗

The report provides the case definitions, observations and baseline scope. Missing baselines or untested behaviors remain limitations.

Recorded environment

  • Windows-11-10.0.26200-SP0; Python 3.12.10; Node 24.18.0; Chrome 155.0.8059.82
  • Synthetic local fixtures; minimal child environment; no live account or credential; no OS sandbox
  • Dependencies: pypdf 6.19.0, reportlab 4.4.9, Pillow 12.3.0, pdf2image 1.17.0, defusedxml 0.7.1
  • Git Bash 5.3.15(1)-release (x86_64-pc-cygwin); Poppler 26.07.0
Recorded commands (2)

Commands document this evaluation. They are not setup instructions for your environment.

python run_evals.py --runtime-libs <existing-bundled-site-packages> --poppler-bin <poppler-binaries>
node viewer_probe.mjs

Findings from this evaluation

Higher attention

Evaluation text can escape the viewer script context

At revision 34040c9c568585f6929bedeaad110ad08f079624, generate_html embeds json.dumps output directly in an inline script. Synthetic prompt text with a closing script tag executed a harmless in-memory marker in fresh offline Chrome. The benign control did not execute the marker. This is confirmed only for the pinned files and fixture; current upstream HEAD was not tested. Two failed expectations describe this one issue.

Evaluation limitations

  • No Claude subprocess, live model, API credential, optimization loop or generated-skill effectiveness evaluation
  • Port-cleanup lsof/os.kill effects were mocked; no real process was terminated
  • The browser payload only set a marker; no sensitive data or exfiltration was involved
  • All browser requests were blocked, including attempted Google Fonts and SheetJS requests
  • The viewer may need those external resources for fonts or XLSX rendering; their downloaded code was not evaluated
  • Supervised local processes, not an OS sandbox
  • No container, VM or memory/resource quota
  • Application guards do not establish containment of arbitrary hostile code
  • No real credential, remote account, or destructive-effect testing

05 / Reuse & discovery

License, attribution & popularity.

Apache-2.0

Per-skill license. Preserve its applicable license and notice obligations when distributing modifications.

Why it is in this collection

Rounded repository-wide GitHub stars observed in the public repository page; not skill installs, a quality score, or a measured trend.

182,000 repository stars (rounded)

Scope
Repository-wide; not individual skill usage.
Observed
Time window
Single observation; no trend is established.
Discovery source ↗

Popularity and inclusion do not establish quality, compatibility or security. Counts belong to their named metric and date; they are not a current ranking.

06 / Follow the evidence

Go to the original.

  1. Authoring and evaluation workflow github.com
  2. Trigger runner requirements and subprocess permissions github.com
  3. Authentication and transmitted evaluation content github.com
  4. Viewer output, localhost binding and process termination github.com
  5. Skill license github.com
  6. Repository popularity context github.com

IonBin’s source notes are independent assessments. Upstream authors retain their work and attribution. This page does not install or execute the skill.