Tatiana Afanasyeva

Tatiana Afanasyeva

VP of Marketing at Bubble

About

Marketing and revenue leader with 15+ years of experience launching and scaling global technology brands, achieving sustainable revenue growth and building high-performing marketing teams. Currently, I lead marketing at Bubble and an active advisor to early-stage startups on post-PMF revenue scale (1-10+MM).

Badges

Tastemaker
Tastemaker
Gone streaking
Gone streaking

Forums

Changing one line at a time saves hours of guesswork.

Two weeks ago, I changed three lines of a system prompt at once to fix tone and formatting. The output quality got worse. I had no idea which change caused the drop. That was a waste of time.

Now I keep a changelog for every prompt edit. I test each change on ten examples. I change one instruction per round. I only keep the version if the accuracy improves.

After three rounds, our draft approval rate went from 62% to 81%. We stopped spending afternoons rewriting prompts from scratch.

Small, isolated edits show you what actually works. Think of changes like code commits each one tested, each one clear.

•

2d ago

Instagram shows a retention graph but gives no data. I had Claude read the graph’s pixels.

Instagram s insights show a retention graph for each reel, with no export and no table. I wanted real numbers to compare reels, so I took four screenshots and asked Claude to extract them.

How it worked: the line on each graph is one distinctive color. Claude found the pixels of that color in every column of the screenshot. It used the 100%, 50% and 0% gridlines to turn pixel height into percent, and the time axis to turn horizontal position into seconds. The output was a list of time and percent points per reel, with a stated error of about 2 percentage points.

What it showed (approximate): a 13-second reel kept about 63% of viewers at 5 seconds. A 35-second reel kept about 44%, and half were gone by 4 seconds. A 30-second reel kept about 70%. A 56-second reel kept about 59%, held around 35% for most of its length, then fell to about 14% in the last 2 seconds.

The weak spots. A screenshot is compressed, and the line is several pixels thick with soft edges, so 2 points is a tolerance, not a measurement against the true data. The method only works when each chart has one line and visible gridlines. And four reels is a tiny sample. It shows a pattern on my account, not a rule.

View more