How HavenSkills reviews Agent Skills
See how HavenSkills reviews Agent Skills source, documents install requirements and records reproducible PDF, Word and Excel file checks.
An install command is only one part of choosing a skill. A useful review tells you when the skill applies, what it needs to run and what a published check actually verified. Alex Che maintains the review criteria for HavenSkills.
Source review
We look for the original publisher and open the skill’s SKILL.md. Its frontmatter should name the skill and describe the tasks that should trigger it. We read the instructions and any files they point to, then record relevant prerequisites, such as a separate executable, a workspace permission or a file format. Where a guide gives a reviewed commit, its Source repo link opens that exact source. An installation command can fetch the repository’s current version.
Recorded file checks
These checks used synthetic files and report the particular output property that was verified. They cover individual file operations.
- Anthropic PDF — Reopened a reordered two-page PDF. The A4 and Letter pages remained in the expected order in the saved file. Scope: page order and dimensions.
- Anthropic Word — Reopened a DOCX and confirmed that its title style and two table headers survived. Scope: saved document structure.
- Anthropic Excel — Reopened an XLSX and confirmed that two inputs and the saved SUM formula survived. Scope: saved spreadsheet inputs and formulas.
Each linked record includes the tool, test date, observation and scope.
How we would compare two agent runs
Use the same verbatim prompt, input files, client and model version for both runs. Keep the workspace and permissions equivalent. Run once with the skill available and once without it. Define an acceptance check before starting, such as the preserved row count or a passing test. Save the output, any failure, elapsed time and token usage where the client exposes it. Record the skill revision and whether the client loaded it.
An example prompt on a guide is a starting point for a check in your own client. Published check results appear in the guide’s HavenSkills workflow check section.