edit mode, and instrumenting the truth

the headline this month is edit mode. the idea is simple: most of the time you already know what you want to say, you just need it to land. so you paste a tweet you wrote, and you get back a batch of versions that keep your idea, your references, and your voice, and only fix the delivery. tighter, the punchline moved to the end, the windup deleted. it runs on the same model as create, and it's built to respect the original instead of turning it into a different tweet.

edit also lives on every draft you generate. there's an edit button on each tweet now: pick tighter, punchier, shorter, or sharper, or just type what you want changed. one credit, and it swaps the tweet in place. it replaced the old regenerate-with-feedback flow, which did roughly the same thing with more buttons.

the other big change is that i stopped trusting my own metrics. i was only counting clicks on my own copy and download buttons, which miss the most common behaviour: selecting the text and hitting cmd+c, or long-pressing an image to save it. those are tracked now. i also added latency timing and fixed the pricing page-view event, which had never once fired.

i spent real time in the data this month. the most useful finding: people who try the product once and leave are not typing lazy inputs, they're pasting substantial stuff. shitpost output is genuinely good. but sensible output kept reaching for the same handful of structures, so it felt formulaic. i banned the worst offenders in the prompt so each tweet has to find its own shape.

i also cleaned up the tweet card (redundant buttons gone, save made consistent, copy promoted to the primary action), archived the remix and visuals modes that almost nobody used, and did a full copy audit so the landing page, pricing, and docs all describe what the product actually is now.

the theme for next month is depth: turning a first generation into a habit, and making the first batch land more often.