One Thing AI Collaboration Taught Me Today
Working on real projects with AI agents like Claude and Codex today, I hit a very basic problem again: the more I rush the AI for an answer, the more likely it is to give me a plan that looks complete but might not actually work in practice.
It reminded me of a principle I keep coming back to: the key to working with AI isn't asking cleverer questions. It's defining the problem more clearly.
Start by stating the goal in plain language
I work full-time on my own products and content now, and I also take on custom software projects using AI. What I deal with every day isn't demo-level stuff. It's real systems: scraping, scoring, summarizing, distribution, pages, automation, data flows.
So I can't just tell the AI "help me optimize this." First I spell out which part of the workflow we're fixing, what's actually hurting the user or the business, and what a good result looks like.
A lot of the time, just getting the goal clear solves half the problem.
Boundaries matter more than prompts
I used to obsess over writing smarter prompts. Over time I realized that on real projects, boundaries matter more.
For example:
- Which files it can touch, and which configs are off-limits
- No leaking production credentials
- No making up customers, revenue, or partners
- When a real system is involved, it has to respect the existing architecture
Not exactly exciting, but it works. The AI is perfectly capable. It just doesn't know where my red lines are.
Have the AI play it back before it starts
These days my usual flow is to have the AI restate the problem and its plan in its own words first.
- If it gets something wrong, I correct it right away
- If it missed a constraint, I add it
- If it's scoped the problem too broadly, I narrow it back down
This step feels slow, but it saves a lot of rework later. As a solo developer, slow code isn't what scares me. Going a long way in the wrong direction is.
Verify in small steps, don't expect it right the first time
AI helps me write code faster, track down bugs, and sort out my thinking. But in the end it all comes back to verification: do the tests run, does the build pass, does the page actually work, does the logic fit the business?
I'd rather treat AI as a really strong collaborator than an automatic delivery machine.
Today's takeaway is simple: AI speeds things up, I make the calls. Once goal, boundaries, playback, and verification become routine, AI can really help one person turn an idea into a product.
If you're also building real projects with AI, drop a comment: where do you get stuck most often?