Using Claude to Write Instructions for Claude

There's something delightfully recursive about asking an AI to help you get better results from that same AI. It sounds like a philosophical puzzle, but it's actually one of the most practical things I've started doing with Claude.

The idea is simple: Claude is remarkably good at researching best practices and synthesising them into structured guidance. And Claude is also remarkably good at following structured guidance. So why not have it create its own instructions?

The Workflow

I'll walk through this using bug reports as an example, because it's something I do all the time and the quality difference and productivity gains are genuinely noticeable.

Step one: research. Ask Claude to research what makes a good bug report. Don't give it constraints yet, just let it pull together what it knows about the topic. It'll come back with something comprehensive covering things like reproduction steps, expected versus actual behaviour, environment details, severity classifications, and so on.

Step two: convert to instructions. Here's where it gets interesting. Take that research and ask Claude to turn it into instructions that would guide an LLM in creating a bug report. You're essentially asking it to write a prompt for itself. The result is a structured template with clear guidance on what information to gather and how to format the output.

Step three: create a project. In Claude, create a project specifically for Jira tickets (or whatever your issue tracker is). Copy those instructions into the project knowledge with a name like "Instructions for creating a bug report". Now every conversation in that project has access to this guidance.

Step four: use it. When you need to create a bug report, start a conversation in that project and ask Claude to help. It will follow the instructions you've established.

The Key Detail That Makes This Work

Here's the thing that transforms this from a basic template into something genuinely useful: include something like "ask me clarifying questions to extract the details for the bug report" in your instructions.

This single addition changes the interaction completely. Instead of you trying to remember everything upfront, Claude will guide you through a short conversation to pull out the details. What were you trying to do? What happened instead? Can you reproduce it consistently? What browser and environment?

You end up with bug reports that have all of the details your developers actually need to solve the problem, without you having to remember a mental checklist every time.

A Shameful Observation

Before I started doing this, I'd notice a thousand tiny details across all of the products I worked on. Small inconsistencies, minor bugs, little friction points. The volume was so overwhelming that most of them never got reported. The cognitive overhead of writing up each one properly meant they just accumulated in my head or got lost entirely.

Now, with this workflow, it's a lot easier to actually capture them. When I notice something, usually on a Teams call, I make a quick reminder and create the bug report later. And the result is that the products are genuinely getting better because of it. Issues that would have lived rent-free in my brain for months are now documented and fixable.

Taking It Further

If you're using the Jira MCP integration, you can have Claude post the bug report directly to Jira for you. I'll write more about MCP integrations at some point, but the short version is that Claude can interact with external tools when you enable these connections.

The final step, which still needs a human touch, is adding screenshots or screen recordings to visually illustrate the problem. Some bugs are nearly impossible to describe in text but immediately obvious when you see them.

Why This Matters

The broader principle here isn't really about bug reports. It's about recognising that Claude can help you create the scaffolding for better interactions with Claude. Any recurring task where quality matters and consistency is valuable is a candidate for this approach.

Design reviews. User interview synthesis. Meeting notes. Technical specifications. Anywhere you find yourself wishing you had a template but never quite getting around to making one.

The meta quality of this workflow, using an AI to improve your use of that AI, feels a bit strange at first. But it's genuinely one of the most practical things I've done to get better results from these tools.