From a Week to a Morning: My AI-SPEC Ice Fishing Production Line

The Morning That Changed My Ice Fishing Workflow

Last Tuesday, I sat at my desk with five ice fishing gear requirements staring back at me. Each one represented about 32 hours of development work. Normally, that's a full week of grinding: clarifying business rules, drafting PRDs, sketching prototypes, then waiting for the team to poke holes in everything.

Instead, I had all five specs done by lunch. No all-nighters. No frantic Slack pings. Just a system I've been tuning for months—a knowledge-base-plus-three-expert-agents pipeline that handles the heavy lifting. It didn't happen because I found some magic prompt or switched to a smarter model. It happened because I stopped trying to out-prompt the AI and started feeding it the right stuff.

Why Your Prompt Isn't the Problem

Most people I talk to think the key to getting good AI output lies in crafting the perfect prompt. They tweak wording, stack constraints, add more rules—and still get generic, off-the-mark results. I've been there. The real issue isn't the prompt; it's the context.

Large language models already have solid logic and writing skills. What they lack is your domain knowledge. If you don't feed them your specific ice fishing specs—like auger blade diameters, ice thickness charts, or the exact torque settings for your drill clutch—they'll just invent something plausible but useless. My mantra is simple: good stuff in, good stuff out. Build a high-quality knowledge base first, then let the AI use it. That beats any prompt tweak.

Feeding the Beast: From Scraps to Structured Markdown

My first attempts at building a knowledge base were a mess. I threw in scanned PDFs and photos of old spec sheets. The OCR mangled numbers, field names got garbled, and the AI would confidently read a 10-inch auger as 10 meters. Then I tried Word docs and spreadsheets—better, but still cluttered with formatting that confused the model.

Finally, I settled on pure Markdown. Clean structure, no noise, clear hierarchy—it's the ideal format for AI parsing. But I don't just dump files in. I break down historical ice fishing requirements and operations manuals into individual Markdown docs by subdomain. Then I have the AI do a second pass: standardize terminology, extract core rules, map out edge cases. Only after that do I import them into Dify—a tool that handles segmentation, retrieval tagging, and API access. It turns the docs into searchable fragments, so the AI pulls exactly what it needs without wasting tokens on irrelevant info.

Here's the thing: a knowledge base isn't a one-time setup. It's a living asset. Every time a new ice fishing spec ships, I extract the changes and merge them into the relevant subdomain. I keep timestamps so the knowledge always matches the current reality. And when the AI gives a wrong answer, I don't rewrite the prompt. I ask it which knowledge chunk it used, then fix the source.

Three Experts, One Production Line

With the knowledge base in place, I designed a workflow with three specialized agents instead of one mega-agent. Each mirrors a real role in my product team.

1. The Research and Planning Expert

This is the "do the right thing" stage. It reads the knowledge base, brainstorms around my initial idea, asks clarifying questions, and helps define scope and business goals. It doesn't decide anything—that's my job. But it makes sure we've covered angles I might have missed. It also flags if a requirement conflicts with existing gear specs or seasonal constraints.

2. The Design Expert

Once the framework is set, this agent takes over and writes the actual SPEC. It follows a strict format: core logic, workflows, boundary conditions, acceptance criteria. No fluff, no storytelling. It also generates an HTML prototype—not embedded in the doc, but as a separate file the AI can read directly. This is where the time savings really show. What used to take me a full day now takes minutes.

3. The Risk Review Expert

Last but not least, this agent plays the QA/reviewer. It checks the SPEC and prototype for format violations, logical gaps, unclosed scenarios, and potential business risks. If something's off, it flags it for human review. This step used to involve dragging developers and testers into a meeting to hunt for issues. Now the AI catches the obvious stuff first, and I only have to deal with the genuinely debatable points.

Why I Ditched Superpowers for a Custom Setup

You might wonder: why not use an all-in-one tool like Superpowers? I've tried it. It's powerful, no doubt. But it's built for full-stack developers—it covers everything from idea to code, including testing and deployment. As a product manager, I only need the front half: from idea to SPEC. The extra steps just burn tokens and slow things down.

My custom framework is lighter, faster, and more flexible. I can run multiple instances in parallel, which is how I knocked out five requirements in one morning. It's not about which tool is objectively better; it's about matching the tool to the job.

When the Knowledge Base Doesn't Help

Let's be honest: this method has limits. It shines in mature product lines where you have years of accumulated specs and history. But if you're designing a brand-new ice fishing gadget—say, a heated ice scoop—there's no existing knowledge to draw from. The old data might actually box you into conventional thinking.

In those cases, I switch to a different mode. I add a research agent that pulls industry reports, competitor info, and any new business context I provide. This builds a fresh foundation before the normal pipeline kicks in. The rule: rely on historical knowledge for existing products, rely on research for new ones. Don't force one approach everywhere.

What Actually Matters in the Age of AI

That morning wasn't magic. It was the result of a solid knowledge base and a well-designed workflow. I'm convinced that the product manager's core value is shifting. It's no longer about writing beautiful docs or fancy prototypes—AI will do that faster and more consistently. The real differentiators are two things: the ability to turn scattered business experience into structured, AI-readable knowledge assets, and the ability to design collaborative workflows that let AI handle the grunt work.

AI isn't just a tool that writes for you. It's a team you can orchestrate. Your job is to be the manager.