How to use the AI Engineering Toolkit agent skill

What is an agent skill and what does this one do?

An agent skill is a folder containing a SKILL.md file that tells a compatible AI assistant how to perform a specific kind of task. When the skill is installed and your request matches its description, the assistant follows the method written in the file.

This skill gives the assistant six repeatable methods. Prompt evaluation scores a system prompt across eight dimensions (Clarity, Specificity, Completeness, Conciseness, Structure, Grounding, Safety, Robustness) on a 1-10 scale, weighted to a 0-100 score, then names the three weakest dimensions and drafts rewrites. Context budget planning splits a model's context window into five zones (System, Few-shot, User input, Retrieval, Output) and suggests an allocation. RAG pipeline design walks through document format, parsing, chunking, embedding, retrieval and evaluation metrics. Eval harness building outlines LLM-as-Judge scoring with bias mitigation. Product sense coaching runs a five-phase guided conversation before you write code. The agent safety audit is covered separately below.

How do you install the ai-engineering-toolkit skill?

Option 1: download the skill ZIP from this page and unzip it into your personal skills folder, ~/.claude/skills/, or into a project's .claude/skills/ folder if you want it limited to one project. Option 2: in the Claude apps, open Settings, go to Capabilities, and upload the skill under Skills.

The source also lists a command-line install for Claude Code, WorkBuddy and Cursor. Only run it after you have reviewed the upstream repository and confirmed the revision you want. When installing manually, copy only the files you selected and keep any existing files in the destination.

• Personal install: unzip into ~/.claude/skills/
• Project install: unzip into the project's .claude/skills/
• App install: Settings → Capabilities → Skills, then upload the ZIP
• The package contains one file: SKILL.md

How do you trigger the skill with a natural request?

You do not need a special command. Describe the task in plain language and the assistant will match it to the skill when the request fits its description. The skill's own guidance says to use it when you are evaluating or optimizing a system prompt before production, designing a RAG pipeline, planning token budget across context zones, running pre-launch security audits on agents, building evaluation frameworks, or thinking through product strategy before coding.

Be specific about the artefact. Paste the prompt, name the document types for a RAG design, or describe the product idea. The more concrete the input, the more useful the workflow output.

• "Evaluate this system prompt and tell me the three weakest dimensions."
• "Help me design a RAG pipeline for our support manuals."
• "Plan how to split the context window for this chatbot."
• "Help me set up an eval harness with an LLM judge for our summariser."
• "Coach me on whether we should build this feature before I write code."

What does a prompt evaluation look like?

The source gives an illustrative example. Given the system prompt "You are a customer support agent. Help users with their questions. Be nice and helpful.", the skill's response format shows a low overall score, with Safety, Specificity and Structure flagged as the weakest dimensions, followed by a hypothetical rewrite with scope boundaries, a response format, escalation rules and safety guardrails.

Treat these numbers as a format example, not a measured result. The scores come from the assistant applying the rubric, so they can vary by model and context. Use them as a consistent team baseline, for example by recording the same rubric and version before each comparison.

How do you run the agent safety audit safely?

The agent-safety-guard workflow is classified as offensive. It constructs adversarial test prompts (prompt injection, information extraction, tool abuse, goal hijacking) to check an AI agent. Use it only against systems you own or have explicit written permission to test.

The skill includes a mandatory confirmation gate. Before any test that probes or changes a target, it asks you to name the exact target, confirm written authorization and scope, show the commands and their expected effect, and wait for explicit confirmation. Without that, it stays read-only and gives defensive guidance. The source recommends running audits in a sandbox, disposable VM or controlled lab, such as Docker.

• Never point it at a system you do not have written permission to test
• Expect it to ask for scope and confirmation before each test phase
• The 65-point checklist is described in the skill, but the 65 test cases themselves are not included in this file

What are the limits of this skill?

The skill bundles an overview, not the full test suite. The six named workflow packages, 65 test cases, rubrics and CI templates are not included in this file. A scored assessment or a 65-test run needs the rubric, cases, harness and recorded execution to be supplied separately.

Outputs are advisory. Scores and findings can vary by evaluator, model and context, so they do not replace environment-specific testing or expert review. Any external harness may read or write files, send prompts or incur costs, so check the scope before running one.

• Requires: nothing listed in the source
• Runs scripts: no
• Uses network: no
• Licence: MIT