I Turned an Anthropic Prompting Tutorial Into an Open-Source AI-Agent Skill

Recently I watched a public Anthropic tutorial about writing better prompts for LLM applications:

https://www.youtube.com/watch?v=ysPbXH0LpIE

The tutorial walked through a practical example: start with a simple prompt, watch the model fail in a specific way, add more context and structure, run it again, and keep improving the prompt until the model can reliably handle the task.

That sounds simple, but the lesson is deeper than “write clearer instructions.”

The real lesson I took away is this:

A production prompt is not just a paragraph. It is a harness.

It defines the model’s job. It gives the model stable context. It separates changing input from reusable reference material. It tells the model what order to inspect information in. It defines how uncertainty should be handled. It shapes the final output so an application can use it.

After watching the tutorial, I wanted to turn those ideas into something reusable.

So I built an open-source AI-agent skill:

Prompt Harness Architect

Repo:

https://github.com/evanwhl508/ai-agent-skills

Why I Made This

When people talk about prompt engineering, it can sound like the goal is to find the perfect sentence.

But in real applications, a prompt usually has more moving parts:

• What is the model supposed to do?
• Who is it helping?
• What information is stable across requests?
• What input changes every time?
• What should the model inspect first?
• What should it do when the evidence is unclear?
• Should it return prose, JSON, XML, or some mixed format?
• How do we know the prompt actually works?

The Anthropic tutorial made this very concrete. The prompt improved not because of one magic phrase, but because the prompt became more like a structured workflow.

That is exactly the kind of behavior I want from coding agents when I ask them to help me build prompts.

I do not want the agent to only rewrite my prompt in nicer English.

I want it to help me design the harness around the prompt.

What the Skill Does

prompt-harness-architect helps an agent design or improve prompts for production LLM workflows.

It guides the agent to think through five phases:

1. Contract Define the task, role, success condition, failure modes, and uncertainty behavior.
2. Structure Separate stable context from dynamic input. Put schemas, rubrics, policies, artifact layouts, examples, and domain rules where the model can reliably use them.
3. Steer Add the procedure the task needs: what to inspect first, what to compare, when to abstain, and when to use examples or conversation history.
4. Format Decide whether the output should be human-readable text, strict JSON, XML tags, or a mixed response. Use response prefill when the runtime supports it.
5. Evaluate Create eval cases for happy paths, missing information, ambiguity, conflicting evidence, unreadable input, output-format compliance, and prior failures.

The skill also includes references for:

• prompt harness templates
• eval case templates
• runtime notes across different harnesses
• source and attribution notes

Why Make It a Skill?

I am interested in building open-source skills that make AI agents better at real work.

A skill is useful because it packages a repeatable workflow into something an agent can load when needed.

Instead of explaining the same prompt-engineering principles every time, I can install the skill and ask the agent to apply it.

For example:

npx skills add evanwhl508/ai-agent-skills --skill prompt-harness-architect

Then I can ask an agent:

Improve this chatbot system prompt using the Prompt Harness Architect skill.

Or:

Design a prompt harness for document/image review and include eval cases.

The agent should then think beyond surface wording. It should consider context structure, uncertainty, examples, output formatting, and tests.

Portable Across Agent Harnesses

I also wanted the repo to be portable.

The canonical skill lives under:

skills/prompt-harness-architect/SKILL.md

The repo also includes thin adapter metadata for different harnesses, including Codex and Claude Code, plus planned notes for Cursor and OpenCode-style workflows.

The idea is:

Author once. Adapt at the edge.

The skill itself should stay harness-neutral. Any tool-specific metadata should live around the skill, not inside the core instructions.

What I Learned

Building this made me appreciate prompt engineering as an engineering discipline.

The important loop is not:

write prompt → hope it works

It is:

write a simple prompt → observe a concrete failure → add missing context or structure → rerun → turn the failure into an eval

That loop is what I wanted the skill to capture.

Try It

This is v0.1.0, and it is intentionally small. I plan to add more skills over time.

You can try it here:

https://github.com/evanwhl508/ai-agent-skills

Install:

npx skills add evanwhl508/ai-agent-skills --skill prompt-harness-architect

Feedback, issues, and suggestions are welcome.

This project is inspired by public prompting guidance and my own attempt to turn that guidance into a reusable AI-agent workflow. It is not affiliated with Anthropic.s