Agent Skill Readiness: A Frozen Metadata Snapshot

The denominator changes the finding.

The source report exposes 1,000 quality-ranked records, including tools, workflows, collections, and generic projects. Only 97 have the source classification agent-skill. We retained every row and analyzed that narrower cohort separately.

Across the full sample, 32/1,000 (3.2%) have an install-path flag. Within the agent-skill subset, the figure is 29/97 (29.9%). Applying the first percentage to agent skills alone would mix unlike records.

In that subset, 89/97 (91.8%) have a recorded license value, but only 28/97 (28.9%) combine a recorded license, a timestamp within 180 days, and an install-path flag. Those are metadata observations, not results from running the Skills.

Two cohorts, explicit denominators.

The 1,000 records reference 978 distinct repository strings. The 97 agent-skill records reference 78. Several records can belong to one repository; these counts measure records, not independent projects.

Inside the agent-skill subset, 28 records have both a recorded license and an install-path flag; 61 have a license value without that flag; 1 has the flag without a recorded license; and 7 have neither. The CSV preserves the underlying field values so each observation can be checked.

Methods and boundaries

Selection. The highest-ranked 1,000 approved candidates by the current quality ordering. The sample is a product audit, not a random statistical sample. This analysis uses the complete skills array, not the 25-entry topSkills shortlist. Classification is copied from the source; we did not manually validate the classifier.

License value recorded

A non-empty license value outside the unknown-value list. License texts and permissions were not reviewed.

Timestamp within 180 days

The exported timestamp is within 180 days of source generation. It can reflect a registry edit rather than a repository push.

Install-path flag present

The registry records a primary install command. Instructions, dependencies, compatibility, and installation success were not tested.

All three signals on one record

The intersection of these three metadata observations. This is not a readiness certification.

License rule. After trimming and lowercasing, blank, unknown, none, n/a, and unlicensed are treated as not recorded. Other values remain metadata labels; they do not establish permissions or compatibility.

Time rule. The exported timestamp must be between zero and exactly 180 days before source generation. The reference time is fixed at 2026-10-18T14:50:46.144Z. Missing or invalid dates and future dates remain separate. There are 0 unknown dates and 0 future dates in this capture.

Installation instructions cannot be assessed from this export. The JSON contains a boolean, not the README, installation command, prerequisite list, or execution log. This report deliberately measures that boolean's coverage instead of claiming to test instructions.

• A quality-ranked, non-random registry sample; percentages do not estimate the wider Agent Skill ecosystem.
• Source classification is a metadata heuristic, not a manual confirmation that a repository contains a usable Skill.
• installReady is copied as installPathFlag: the registry can produce a primary install command. It does not measure README instructions, prerequisites, compatibility, installation success, or task success.
• The source lastPushedAt field may fall back from the GitHub push timestamp to the registry updated_at timestamp. Recent metadata cannot establish recent code maintenance.
• License values are copied from registry metadata; license texts, SPDX validity, permissions, and legal compatibility were not inspected.
• Multiple records may refer to one repository. Record-level observations are not independent repository-level observations.
• No third-party Skill was downloaded or executed. Trust scores and reported agent outcomes are not evaluated in this study.
• The upstream report can use a curated fallback when its live query fails; its JSON does not expose which read path ran. This study preserves the public response without claiming direct database provenance.

The reviewed source definitions are pinned to repository revision be2b8d6: the report generator and supply profile. The API response does not identify its deployed code revision; the preserved response below is the observed evidence.

Download the evidence and reproduce it offline.

The raw source was fetched once from the public report JSON. Its exact bytes are preserved alongside the capture time and SHA-256 digest. Reproduction makes no network requests and runs no third-party Skill.

• Frozen source JSON
• Capture record and SHA-256 hashes
• Analysis and row-level JSON
• Row-level CSV
• Reusable SVG chart
• Full methods in Markdown
• Standalone Node.js script

Save all seven files in one folder, then run this command with Node.js 20 or later:

node reproduce.mjs --check

The script verifies the frozen source hash, regenerates the derived files in memory, and fails if any output differs. Dates do not move when it is rerun.

Cite this version.

OpenAgentSkill (2026). Agent Skill Readiness: A Frozen Metadata Snapshot. Version 1, captured 2026-10-18. https://www.openagentskill.com/reports/skill-readiness-2026-10-18

When quoting a percentage, include the cohort and denominator—for example, “29 of 97 source-classified agent-skill records in OpenAgentSkill's quality-ranked snapshot had an install-path flag.” This sample does not support a claim about all Agent Skills.

For this site's original code, chart, and analysis, see the repository's existing MIT license. This report does not grant new rights to the upstream raw dataset or third-party software. Please cite the dated report when quoting its findings; a citation is not an endorsement or a verification badge.

For the live registry context, see State of Agent Skills 2026. Maintainers can use the Creator Kit to document a real case and supply clearer installation information.