Prompt Seagulls

I was listening to @syntaxfm this morning and @wesbos said at 328s “The prompt seagulls are driving me nutty”. 100% hard agree Wes. Now I didnt just come off of a conference where they dropped fable 5. I’ve not logged into twitter in months, but I see it around work.

There is no such thing as magic #

Asking for the prompt as a first question here, is a clear sign you are looking for magic. What are the magical incantations I must pass to all master clanker to get it to do the things that I want. Wrong, there is no magic prompt to get impressive results, its people that know and understand their craft that get amazing results. They have taste. They have design in mind, They see what the end product should be. They are guiding the model to do what they want and are not surprised when they get it.

process #

No one is one shotting their work and kicking up their feet for the other 39 3/4 hours of the work week. Theres process and methodology to it. Develop the right skill files, the right AGENTS.md, instructions, all of the other things. This puts the ball in your court telling the models how you want these things done without deep prompting.

taste #

If you want to build products that humans use and enjoy you must bring some level of domain expertise. You should know a bit about your product, your customer, and how the underlying tech you are working with works. You should understand the basics of how websites are built and laid out or how a cli operates. If you don’t you will inevitably build out something that does the job, but is not intuitive, does not look good, or feel good, and no one can figure out how to use it, cause no one, especially these days, no one is reading the docs.

testing #

Actually try to use your product, can you? Can you hit errors? You probably should be able to one way or another if you are accepting input from a user they are bound to be able to input something you cannot just automatically recover from. In these cases does it make sense, can you easily see what you were supposed to do, or why it was wrong? Does it work on various resolutions? This goes for about everything these days, web or terminal. Can you read the output on a full screen easily? What about a half screen? If you’re a Philz coffee mac fanboi with slinging a top of the line 8k display, does it even work on what a mere mortal might be trying to use?

Good starting point #

Ai repeats patterns it finds, its very good at repeating patterns it finds nearby. If your existing codebase is not consistent, and riddled with pitfalls, it will be replicated and be digging larger holes for you to dig up later.

Old School Checks #

Automated checks are a hard requirement for me these days, lint, type check, secret scan, cve scanning. All of these things help drive consistency and helps reliability.

Actually Read Some of the code #

Now for the most controversial, un-vibe coder thing… Actually try to read the code. You might notice some things. Somethings like regression testing, and supporting old versions, old formats, duplicating schemas. AI is really good at waking up from a cold state, assuming everything it sees is in production with real users and writing a lot of code. It will do things like supporting the version that it made 5 minutes ago and never shipped for eternity. Have your clankers review too, that’s also fine, not a substitute for you taking a look. Even if you are not finding the tiny bugs here, you will understand more about how your project works, the direction its going, and be able to catch directionally wrong shifts.