How to share what you built (without it sounding like a pitch)
A short, practical guide to posting your AI work so people actually read it, reply to it, and learn something from it.
You built something with AI. You want to share it. And then you open the composer and freeze, because everything you type sounds either like a brag or like a press release.
Here is the format that works, and it takes about four minutes.
Start with the result, not the setup
Most people write three paragraphs of context before getting to the point. Flip it. Open with what changed:
- "Appointment bookings went from 41% to 63%."
- "A 60-second ad, every shot AI, made in a weekend."
- "Cut our transcript costs by 80% by switching one model."
The context can come second. If the first line doesn't tell someone whether this is relevant to them, most of them will not reach the second.
Name the tools
"I used AI to build this" tells nobody anything. "Vapi for the call, Claude for the extraction, Supabase for storage" tells people whether your experience transfers to their stack. It also means your post shows up when someone searches those tools six months from now.
On AfterPrompt the tool tags are part of the post for exactly this reason.
Say how long it took
This is the single most useful number you can include, and almost nobody does. "Two evenings" or "three weeks, most of it on auth" tells people more about feasibility than any feature list.
Then the two boxes
- What worked: the specific decision that made the difference
- What didn't: the thing you tried first that failed
Be concrete. "Prompt engineering" is not an answer. "Putting the examples before the instructions instead of after" is.
What to leave out
- Anything that sounds like a landing page. "Revolutionising how teams..." Delete it.
- Screenshots of a chat window with no explanation.
- The word "excited," unless you actually are, in which case fine.
A worked example
Bad:
Excited to share what we've been building! Our AI-powered platform helps teams unlock productivity. Would love your feedback 🙏
Good:
Our support bot was closing 30% of tickets. Now it's 71%.
The change: instead of one big prompt with all our docs, we retrieve the three most relevant help articles first and pass only those. Same model, a quarter of the tokens.
Worked: chunking by article section, not by character count.
Didn't: letting it answer billing questions. We route those to a human now. It was confidently wrong too often.
Claude, Supabase, about two weeks.
Same project. One of them gets replies.
The bar is lower than you think
You do not need a launch. You do not need numbers. "I spent Saturday trying to get X working and here's where I got stuck" is a great post, and it will get you better answers than asking the model.