← Explore all agents

EVALUATION / Anthropic

Post-hoc skill analyzer

Skill-creator agent that reads a finished blind comparison, then opens both skills and transcripts to say why the winner won.

“Runs after the comparator has chosen A or B, reads winner and loser paths, and writes improvement notes to output_path.”

01 / THE REASONING

Why this made the selection.

  • The catalogue has the skill-creator grader and not this post-hoc analyzer.
  • The file is explicit that it runs only after a blind winner exists, so it is not a second copy of the comparator.

02 / THE REVIEW RECORD

What we actually inspected.

Source review has boundaries.
A clear record is more useful than a “safe” badge.

Material inspected

  • plugins/skill-creator/skills/skill-creator/agents/analyzer.md
  • plugins/skill-creator/LICENSE
  • Round 2: pinned definition, output contract and permission/dependency boundaries reread on 2026-10-14

Our findings

  • The agent receives winner, both skill paths, both transcript paths, the comparator JSON, and an output path.
  • It reads the comparator result before it reads the skills, then looks at instruction specificity, tool use, examples, and edge cases.
  • The file says it unblids the comparison by reading the skills and transcripts after the blind winner is known.
  • The comparison mode examines instruction, tooling, example and reference differences after seeing the winner.
  • The benchmark mode forbids improvement recommendations and instead highlights nondiscriminating assertions, variability and quality/cost patterns.

Not established by this review

  • The agent was not executed.
  • No transcript was read with it.
  • No output file was written.

The review applies to the material and revision named here. A newer upstream release can change its behavior.

03 / PUT IT TO WORK

Use the role in your project.

Upstream setup instructions ↗
  1. No concrete install command without a placeholder is in the marketplace README.
  2. This agent is invoked by the skill-creator evaluation flow, which was not run.
  3. Keep the task mode explicit and retain all run evidence, failed cases and scoring definitions before interpreting an apparent winner.

Before you start

  • A finished blind comparison with paths to both skills and transcripts
  • Claude Code skill-creator

04 / FOLLOW THE EVIDENCE

The source trail.

Our notes are separate from the original resource.
Check upstream before adopting a new version.

  1. analyzer.md ↗Checked

    https://github.com/anthropics/claude-plugins-official/blob/8bd00f08f0b04705e610cf132fbe2cf5ec44cfe3/plugins/skill-creator/skills/skill-creator/agents/analyzer.md

    Supports: post-hoc role and inputs

  2. skill-creator LICENSE ↗Checked

    https://github.com/anthropics/claude-plugins-official/blob/8bd00f08f0b04705e610cf132fbe2cf5ec44cfe3/plugins/skill-creator/LICENSE

    Supports: Apache-2.0

Evidence & Ask