Before
After
For about a week, the Pluck dashboard had a prompt input at the top. You typed "sprint retro" and AI generated the fields. Title, description, six questions, sensible types. It worked. It demoed beautifully. I screenshotted it three times.
I shipped it on a Friday at 11pm. Removed it the following Saturday at noon. A whole weekend building it, the next weekend pulling it back out.
Here is the honest version of what happened. I asked Claude to walk through the dashboard as a brand-new user — pretend you've never seen this product, what do you do first. Claude flagged the prompt input immediately. "This is the agent doing the human's work. The human came here to do the work. Why are you asking them to describe it to an AI on a page they navigated to in order to skip that step?"
I argued back for a while. The prompt input is fast. It's flashy. People like typing things into boxes. Claude was unmoved. I went and made coffee and came back and Claude was still right.
Pluck has a principle I've been trying to hold: same actions, two surfaces — the web is for humans, the AI chat is for agents. A human comes to the dashboard and picks a template. An agent calls the generation endpoint and gets a form back as JSON. Both end up in the same place. Neither does the other's job.
The prompt input on the dashboard violated that. It put an agent-shaped interaction inside the human-shaped surface. And it confused new users in a specific, repeatable way: they'd type "I want to collect customer feedback" into the prompt, get a generic feedback form, then not understand why the templates below also said "customer feedback." Two paths to almost the same thing, neither obviously better than the other.
People asked me, "why would I pick the template if I can just describe it?" And the answer was, well, the template is faster and the output is better-curated. Which is true. So why have the prompt at all? I didn't have a good answer.
The endpoint is still there. Agents use it. If you ask Claude to make a Pluck form, it POSTs to /api/forms/generate, gets back a form spec, and hands you a URL. That is the right surface for that interaction. The dashboard is the wrong one.
I'm not 100% sure the cut was right. Cutting a working feature stings, and I keep a branch around called maybe-bring-back-prompt in case I change my mind. If it comes back, it comes back as an addon, not the front door. A small "describe a form" thing in a corner, not the first thing a new user sees.
The architecture is cleaner now. The dashboard makes sense to someone seeing it for the first time. That has to be worth something. A feature that demos well isn't the same as a feature that fits — I'm still learning the difference.
— Sumit