Sansi, quick question. I've noticed a bad habit in AI-written code: no matter how simple the feature, it walks through every edge case and ends up turning a small function into a small project.
Any way to fix that?
It's because AI was trained on "being thorough." Writing extra code never gets penalized; writing too little does.
From its point of view, more code equals more responsibility.
So if I put "don't over-engineer" in my prompt, does that fix it?
I tried. It doesn't last.
A prompt is just a passing reminder — as the conversation drags on, it slips back into old habits. You can't hold the line that way.
Oh... so what did you do?
I installed a skill called ponytail.
ponytail? Why would it outlast a one-line prompt?
Now that's the right question.
ponytail ships with two hooks that stay resident. Before every turn of work, it re-injects "be lazy first" into the model's context.
Be lazy? How does it actually slack off?
It has a seven-level "laziness scale," and it pushes back question by question before touching any code:
Does this feature really need to exist? Is there an existing component in the repo? Can native browser APIs do the job?
Only when all of those fail does it write code — and only the smallest version that works.
And the results?
The official benchmark: total code volume cut by more than half, and about 20% fewer tokens.
Sounds solid. But you're sounding a little too smooth — there's a catch, right?
The biggest one: at heart it's still a prompt. Long conversations will still drift.
Fine — pros and cons both on the table. I'll install it, and uninstall if it misbehaves.
Good. Practice on small internal tools first, then bring it to key projects once it proves itself.