DATA REPORT · AUGUST 2026

Top 100 Agent Skills in 2026: What the leaderboard actually shows

The public leaderboard is a useful map of attention—not a universal recommendation list. We analyzed its Top 100 entries to separate adoption signals from task fit, source quality, permissions, and risk.

100leaderboard entries
81unique Skill IDs
53.8Mrecorded installs
25official-source entries

WHAT THE DATA SAYS

Four signals worth keeping separate

01

Discovery is the largest single demand signal

find-skills leads the snapshot with more than 3.0 million recorded installs. The strongest need may be finding and evaluating Skills before running them.

02

Development workflows dominate the top ten

Interviewing, architecture review, TDD, browser work, React guidance, and handoffs occupy most of the first ten positions. These are repeatable workflow bottlenecks, not one-off novelty tasks.

03

The leaderboard is concentrated by source family

The largest source rows come from Feishu/Lark, Microsoft Azure, Matt Pocock, and RunComfy. Nineteen Lark Skill IDs appear under both open.feishu.cn and larksuite/cli, so source-aware de-duplication matters.

04

Popularity is evidence—not approval

Install counts do not prove native Agent support, minimum permissions, maintenance quality, runtime behavior, security, or suitability for your repository. Those checks belong beside the rank.

THE LEADING TEN

Top 10 by the captured all-time rank

Open any profile to inspect its task fit, main limitation, first risk, source revision, and Agent-specific installation guidance.

RankSkillSourceRecorded installs

SOURCE CONCENTRATION

Five source rows account for most of the list

This is a count of leaderboard rows, not a judgment of source quality. The Feishu and Lark rows overlap at the Skill-ID level.

  1. 01open.feishu.cn
    26 entries
  2. 02larksuite/cli
    19 entries
  3. 03microsoft/azure-skills
    19 entries
  4. 04mattpocock/skills
    13 entries
  5. 05prime-skills/runcomfy-agent-skills
    11 entries

HOW TO READ THE REPORT

Use rank as a starting signal, then reduce the decision

  1. 01

    Start with the job

    Choose a concrete outcome before choosing a package. A highly installed Skill can still be irrelevant to the current task.

  2. 02

    Check Agent fit

    Separate explicitly native support from a portable SKILL.md path. Portability is useful, but it is not the same as native testing.

  3. 03

    Review access and evidence

    Inspect permissions, source revision, limitations, and runtime-review status before installation.

  4. 04

    Keep the stack small

    Begin with one core Skill and add a guardrail or evidence role only when responsibilities do not overlap.

NEXT STEP

Turn the leaderboard into a practical plan

Browse the complete directory or choose a task, Agent, and priority to build a deterministic 1–3 Skill plan.