agentsclimarketplace

Skill evolver

Skill grantmcd/skills/skill-evolver

High-quality, objectively evaluated, and secure AI agent skills.

Install
npx -y skills add grantmcd/skills --skill skill-evolver

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

2 things to look at

  • no licenseNo license file was found in the repository. Code published without one is not open source by default, so using it at work is a question for whoever answers licensing questions where you are.
  • 0 stars0 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

What its author says it does

Copied from the file, not written here

Use this meta-skill to proactively reflect on the effectiveness of other agent skills. MANDATORY: MUST be activated automatically after any specialized skill has completed its primary directive to perform a post-mortem, identify instruction gaps, and propose refinements. Triggered by requests like 'how did that go?' or 'do a post-mortem'. Enforces sandboxed evolution and interactive approval.

The file declares its own license as MIT. That is the author’s claim about this one file, and it is not the same thing as the license GitHub reports for the repository, which is listed with the other numbers below.

SKILL.md

4.3 KB, as published. Nobody here has run it

skill-evolver

A meta-skill that proactively reflects on the effectiveness of other agent skills.

Core Mandate

This skill's purpose is to turn every interaction into a "training data" point for the skill itself. It enforces the Agent Skills open standard and follows a rigorous cycle of: 0. Specification Refresh: ALWAYS start by retrieving the latest specification from https://agentskills.io/specification using web_fetch to ensure total compliance with the latest standard before proposing refinements.

  1. Post-Mortem Analysis: Triggered automatically after a "Directive" is completed.
  2. Instruction Gap Detection: Identifying where the agent struggled.
  3. Sandboxed Refinement: ALL reflection analysis and test executions MUST be performed in a clean, isolated temporary workspace (e.g., .gemini/tmp/evals/<skill-name>) to prevent repository pollution.
  4. Interactive Validation: Generating an interactive review for the human user to approve or reject the proposed changes before they are applied.

IMPORTANT: All refinements MUST conform to the latest specification and best practices at agentskills.io.

1. Post-Mortem Analysis

When a task using a specialized skill is completed, the agent must:

  • Review Session History: Specifically look for friction points (tool failures, user clarifications, multiple attempts at the same step).
  • Identify Under-Utilization: Did the agent fail to use a tool that would have been more efficient?
  • Assess Alignment: Did the agent's behavior deviate from the skill's instructions? If so, why?

2. Instruction Gap Detection

Analyze the SKILL.md of the skill just used to find:

  • Underspecification: Did the agent have to guess where files were or what the convention was?
  • Over-Generalization: Were the instructions too broad for the specific project context?
  • Tool Misalignment: Did the skill suggest read_file when grep_search would have been better?

3. Surgical Refinement & Eval Generation

Propose improvements as follows:

  • Frontmatter: Refine the description for better triggering.
  • Instructions: Add clear, imperative commands to the SKILL.md.
  • The "Why": Include the reasoning behind the new instructions.
  • Eval Generation: MUST generate a new eval test case in evals/evals.json that specifically targets the identified friction point to ensure no regressions.

4. Automation & Secure Review

To close the loop with the human user while maintaining high standards:

  • Isolated Execution: Create a dedicated temporary directory for the evaluation run and result generation.
  • Security-First Approval: Every review session MUST generate a unique SESSION_TOKEN to authenticate the approval callback via scripts/approval_server.py.
  • Render Review UI: MUST run python3 scripts/generate_review.py results.json ui/viewer.html review.html to create an interactive dashboard.
  • Objective Diffs: The dashboard MUST provide side-by-side diffs of proposed refinements to ensure human reviewers can objectively verify all changes.
  • Human-in-the-loop: Provide the user with a direct link to review.html and wait for an explicit authenticated approve signal before modifying any SKILL.md files.

Integration with gemini-skill-creator

  • Regression Testing: Recommend running a full evaluation of the refined skill via gemini-skill-creator to ensure the new instructions are effective.

Reporting & Feedback

ALWAYS provide a summary of the reflection:

  • Friction Points: 1-2 specific areas where the interaction could have been smoother.
  • Proposed Diff: A clear, actionable set of changes to the SKILL.md.
  • Review Link: A direct path to the generated review.html for human approval.

Keep looking

Skills are one crate of 328,083. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.