Guide · Agent skill
How to use the Agent Performance Improvement Workflow agent skill
Updated · written with AI from the project's own documentation and checked against our catalogue data
This guide explains what the Agent Performance Improvement Workflow skill does, how to install it, and how to ask an AI agent to use it. It is written for readers who have never installed an agent skill before. By the end you will know what data the skill needs, how to trigger it with plain requests, and where human review and your own test tooling come in.
Agent Performance Improvement Workflowsickn33/agentic-awesome-skills · MIT Download the skill Quick start
- 1Download and unzip. Download the ZIP from this page and unzip it so that SKILL.md sits in a skill folder.
- 2Place the folder. Move the folder into ~/.claude/skills/ for personal use, into a project's .claude/skills/ folder, or upload the ZIP under Settings, Capabilities, Skills in the Claude apps.
- 3Gather authorized data. Collect the existing agent's traces, logs or evaluation results that you are allowed to analyze, with private data redacted.
- 4Ask for a baseline. Describe the agent and its problem in plain language and ask for a baseline performance report and failure classification.
- 5Review proposed prompt changes. Check each suggested prompt, example or role change against your own requirements before accepting it.
- 6Test original versus improved. Run both versions in your own harness on the same task set, since the referenced parallel-test-runner is not bundled.
- 7Roll out in stages. Deploy only after human approval and regression testing, and keep a rollback path ready.
What is an agent skill and what does this one do?
An agent skill is a folder containing a SKILL.md file of written instructions that an AI agent can load when a task matches its description. This skill, named agent-orchestration-improve-agent, is a structured playbook for improving an existing AI agent. It is not a finished tool and does not run anything on its own.
The playbook covers four phases: collecting performance data such as completion rate, corrections, retries and tool usage; classifying failure modes by root cause; revising prompts, examples and output formats; and testing original versus improved versions before rolling changes out in controlled stages.
Install the skill from a ZIP file
Download the ZIP file for this skill from this page and unzip it. The skill folder contains SKILL.md. Place that folder in one of these locations, depending on how you work:
- Personal use across projects: ~/.claude/skills/
- Single project: the .claude/skills/ folder inside your project
- Claude apps: open Settings, go to Capabilities, then Skills, and upload the ZIP
Trigger the skill with natural requests
You do not need a special command. Describe the job in plain language that matches the skill's purpose, such as improving an existing agent, analyzing its failures, or running an A/B test on prompt changes. The skill's own guidance says to use it for existing agents with performance problems, failure analysis, structured A/B tests or evaluation suites, and iterative optimization.
Do not expect it to help when you are building a brand-new agent from scratch, or when you have no metrics, feedback or test cases. The skill says it should not be used in those situations.
- "Our support agent keeps using the wrong tool. Analyze its traces and find the failure patterns."
- "Draft prompt changes for our order-lookup agent based on the corrections users made last week, and propose an A/B test plan."
- "Set up a baseline report for this agent using the logs I have attached."
Example workflow from start to finish
A typical session starts with the baseline. The agent gathers the authorized trace window and reports completion rate, corrections per task, tool call efficiency, latency and token use. It then groups failures into categories such as instruction misunderstanding, output format errors, context loss, tool misuse, constraint violations and edge case handling.
Next, the skill proposes prompt changes, such as clearer role definitions, annotated good and bad examples, and critique-and-revise steps. Finally it describes a test plan comparing the original and improved versions, including golden path cases, previously failed tasks, edge cases and adversarial inputs. You review the plan and decide what to run.
What you need before you start
The skill needs real evidence about your agent. Without it, the baseline and failure classification cannot be built. You also need a way to actually run tests. The skill refers to a parallel-test-runner and named agent roles such as context-manager and prompt-engineer, but these are not bundled with the download. Use your own agent harness or evaluation tooling, or report the gap.
- Access to the existing agent's traces, logs or evaluation data
- An agent harness or evaluation tooling that can run the candidate and original versions
- Permission to analyze the trace data, with private information redacted
Safety notes: scripts, network and human review
The skill runs no scripts and uses no network access, according to its listing. It is still a playbook that recommends actions such as deploying prompt changes and configuring recurring reports. The file itself says not to run hosted evaluations, delegate, message teams, or deploy merely because a phase names them. It asks the agent to follow your existing authorization.
Keep a human in the loop for every step. Keep traces private or redacted, cap cost and number of steps, test candidate changes in isolation, and roll back quickly if quality or safety metrics regress. Do not run unattended changes on live systems.
- Do not deploy prompt changes without regression testing
- Review each proposed step before approving it
- Treat the statistics guidance as a starting point and set your own alpha, sample size and design before testing
What people use it for
- Finding why a production agent makes repeated tool selection mistakes
- Turning users' correction patterns into clearer instructions and annotated examples
- Designing a regression test set from previously failed tasks before changing a prompt
- Planning an A/B comparison of an original and a revised agent with predefined success metrics
Frequently asked questions
How do I install the agent-orchestration-improve-agent skill?
Download the ZIP from this page and unzip it. Place the folder in ~/.claude/skills/ for personal use, in a project's .claude/skills/ folder, or upload the ZIP in the Claude apps under Settings, Capabilities, Skills.
Is this skill free and what is its licence?
The skill is listed under the MIT licence in the sickn33/agentic-awesome-skills repository. Check the repository's licence file for the full terms before reuse.
Does the skill run scripts or use the internet?
According to its listing, the skill runs no scripts and uses no network access. It is a set of written instructions for the agent.
Can I use this skill to build a new agent from scratch?
No. The skill says it is not for building a brand-new agent from scratch. It is designed for improving an existing agent with performance data.
Why does the skill mention parallel-test-runner and agent roles that I can't find?
Those helper roles and the parallel-test-runner are referenced by the skill but not bundled with the download. You need your own agent harness or evaluation tooling to run tests.
What should I do if the skill gives generic advice?
Give it concrete evidence such as traces, failure examples, corrections and test cases. The skill says it should not be used when no metrics, feedback or test cases are available.
New agent skills, repos and AI tools — every week
The best new skills and open-source projects, the fastest-growing repos, and one free prompt. Five minutes to read. Unsubscribe anytime.
Our verdict
Use this skill if you maintain an AI agent in production, already have traces or evaluation data, and want a disciplined process for prompt changes and A/B testing. Pick something else if you are building a new agent, have no metrics or test cases, or want an automated tool that runs tests for you, since this skill is written instructions only and requires human approval at each step.
Download the skill