Bryan Collins

Watch my tutorials on and builds YouTube youtube.com/@bryanjcollins

I usually tell AI to ask me questions in a picker so I can steer agents before they build I created a skill called /picker for that

I built Current to help consultants track Irish solar and battery projects. It uses real data, parses official planning records and provides watch lists to consultants. Now, should I launch it?

I built a roadmap engine with Opus 5.5 and Agon. My AI agents get 3 inputs: → DataForSEO MCP for keyword demand → Mobbin MCP for real UX patterns → My GitHub for what’s already built Once a week, it fills a dashboard with build ideas backed by data and working examples, not just AI “insights.” No more staring at a blinking terminal cursor.

I one-shotted this build with GPT-6 Astra It took about an hour. The project uses APIs to help Irish energy and planning consultants find Irish solar and battery projects I dunno if I'll actually build it out, but Astra is lit.

Gotta get that upsell

Before, I spent hours on design research. Now, it only takes a few minutes. Mobbin has a huge library of real app screens and user flows. It also has an MCP. I plugged it into Claude Code and asked for design ideas. It came back with dozens, each with mockups and the real apps doing it. It's a nice addition to any AI build.

Massive AI prompts are off-putting to read, let alone use. You don't need to write endless paragraphs of instructions, which are all the rage on social, to get good outputs. I spent a dozen hours reading documentation from Google and Claude prompt engineers. They offer a simple framework for writing AI prompts. An effective AI prompt has four essential parts. Persona: Tell AI who it should be. This sets the tone and the level of expertise. Task: Specify exactly what you want, with clear parameters. Context:: Share your goals and any relevant examples. Format: Define how you want the output structured. Here's your fill-in-the-blank template to get started: You're a [role] who [key characteristic]. I need [specific task] for [target audience/purpose]. Here's my [context/examples]. Please format the output as [desired format].

Once you have a prompt that works, store it somewhere so you can reuse it in different AI models. I keep reusable prompts for: * Website FAQs * Email subject lines * Brainstorming * Content repurposing * Business planning Here's a template to get you started. Remember, create one file per prompt and give it a descriptive name so you can access it later: > PROMPT NAME: [Descriptive title] > VERSION: [Version number] > PLATFORM: [ChatGPT/Claude/Gemini] > USE CASE: [What it's designed to do] > PROMPT TEXT: [The full prompt] NOTES: [Performance observations, tips for modification] So, where should you store your prompts? You could create a dedicated Claude or GPT project but that's less ideal if you want to use a prompt in more than one LLM. Store them in simple text file on your computer. When you've a few dozen proven prompts, consider a more advanced solution, like Notion or GitHub. I use a structured Git repository for all my AI projects on GitHub.

I deleted thousands of readers from my email list a few days ago. When I started my email list, my mission was to get 1,000 subscribers. I asked friends and family to join my list and gave away a PDF on my website. I hit my goal after a few months. Then, I set a new goal of reaching 10,000 subscribers. I used every strategy I could think of to hit it: Facebook ads, content upgrades, outreach campaigns, competitions, and multiple lead magnets. One lead magnet over-performed. I explained how to turn a fancy Moleskine notebook into a Getting Things Done (a popular productivity technique) system. I took pictures for my guide and wrote a lengthy tutorial. The magnet attracted hundreds of subscribers. But they weren't interested in any of my offers or courses. They wanted the guide only. So, I ended up having to remove them manually from my list. Dozens of hours, a few hundred dollars, wasted. I want to tell you I only made this mistake once. But I also created content upgrades for blog posts. They covered how to get more done in OS X and Excel... things that had NOTHING to do with my core business. Again, I felt good about my rapidly growing subscriber count but less good when they read my emails and were like WTF?! The bill from my email marketing provider was another painful reminder that I needed a strategy, not just another growth tactic. So, I only create lead magnets that align with my core offer these days. I also regularly remove people who are not a good fit for what I write about or sell. My list isn't for people interested in free PDFs, generic productivity tips, or a low-paying content writing gig. I create for serious writers, coaches, and content creators looking to grow their businesses. Growing a list isn't just about hitting an arbitrary or round number on the backend of Kit or Substack. It's more like a dinner party where only the right guests get an invite.

Make Claude run the numbers BEFORE deciding: "Before you give me any conclusion, produce a table with columns: Metric | Value | Command/API call used | Raw output snippet | Confidence (verified/inferred/unknown). Any row you can't fill with a real command goes in a separate 'Unverified — needs checking' list. Do not state a conclusion until the table exists."

Steal this prompt if you want Claude's help fact-checking your work "Use a fresh subagent as an adversarial fact-checker on the report you just wrote. Give it only the report text and the repo/API access — not your reasoning. Its job: list every factual claim, licence assertion, and number; independently verify each with a real command or fetch; flag anything unverifiable or contradicted. Then revise the report and show me the fact-check table alongside it. "

No sh!t

If there's no MPC , I'm not using it

Steal this business idea prompt: Fable took this prompt and created a dashboard of 101+ business ideas for me. It took about 30 minutes. "You're a business strategist. I need you to do a mega analysis of 101 Online projects, products, or SaaS I could build leanly via AI coding and document Criteria are a minimal team, proven success in other markets or countries, a clear monetization path, and not necessarily an oversaturated market Your output should be at least 101 ideas, plus a brief summary of what I can do differently or better, and an example of the working competitors Nice to have, but not Mandatory: Ireland slant [change for your country] Can use a proprietary dataset Aligns with my personal traits Meets VRIO criteria (valuable, realistic, inimitable, organized)"

I like this /insights skill a lot It's like learning AI coding on the job. Plus, it gives me content and ideas to build for the week ahead.

Neat skill I built. It turns long chats into a simple picker so I can answer Claude's q one by one.

Someone ai-coded a replacement for Adobe Creative Cloud. And it's pretty good!

Asking AI ‘does this work?’ is like asking a student to mark their own exam. The answer is usually A+.

“The most valuable part of my trading app isn’t the code. It’s the rules. Credit-only rolls. Smaller positions. Never pay to get out of a trade. I wrote them down once. Now AI enforces them every day, even when I’m tempted not to.”

I coded an AI agent for trading options Dashboards are lit