Pareto 20, No. 9: Cosmin Pop, for proving a prompt works instead of feeling that it does

Eleven prompt variants, 316 cases, a held-out set of hard ones, and a winner that cut the work by almost four times. He writes for test engineers; the method transfers to anyone.

Seat 9 on The Pareto 20 is Cosmin Pop, and it is here for a single discipline most prompt advice skips: measuring.

Who he is

Pop describes himself as a lead test engineer with six years in quality and test automation, ISTQB certified, working in Playwright and TypeScript, and a hardware tinkerer on the side. His blog posts most days. Every AI piece on it carries the subtitle "A QA automation engineer's notes on..." and every one contains code. That is a warning and a recommendation at once: the audience is engineers, and the habits are exactly what a business owner writing prompts is missing.

What to read him for

The A/B test. Prompt A/B testing for an agent task (August 31, 2026) runs eleven versioned prompt variants through 316 cases across 31 sessions, keeps ten deliberately hard cases aside as a held-out set, and scores each variant on how many turns the agent needed and how often its output was valid. Variant 7 averaged 34 turns per task. Variant 11 averaged 9. The difference was three explicit bans on things the agent kept doing, not more encouragement. His line: "There is no positive encouragement to 'think step by step' or 'be thorough.' The problem was always that thoroughness, without constraints, gets expensive and unsafe."

What a wasted token looks like. In the same piece, one habit he banned was reading a whole page's HTML at once: "roughly 22,000 tokens that did nothing." If you have never looked at what your agent reads per step, that number is the reason to start.

Why "did it get the right answer" is not enough. He refuses pass/fail as the main signal "because that is trivially satisfied by emitting pass for everything," and in Testing the author (August 21, 2026) he goes after easy tests: "A skill that gets easy questions right looks great in a demo and adds nothing in production."

Descriptions as routing. Your skill description is a classifier, not a summary (August 20, 2026) is the short version of why vague instructions collide.

What transfers if you never write code

Four habits, none of which need TypeScript: version your prompts and keep the old ones; keep a small set of hard cases you never train on and test every change against it; measure cost (steps, tokens, minutes) alongside correctness; and when the model keeps doing something wrong, ban the behavior explicitly rather than asking it to try harder. He also admits the limit: a constraint tuned to one model "may be redundant, or actively over-constraining, against a future model." Re-run the test when the model changes.

Start here

The A/B test. Skip the code blocks on a first pass; the tables and the reasoning carry the piece.

What he sells, and what that means for you

Nothing, as far as we can find: no course, no newsletter, no affiliate links, no sponsorships. He is employed as a test engineer by a company he does not name. One loose end: a repository on his GitHub is described as made by Currents.dev, a commercial test-analytics company, and that relationship is not explained on his blog.

Not for you if

You will not read past a code block. This is the most technical seat on the list, held for a method rather than a voice. If that is you, take the four habits above and move on.

---

Part of The Pareto 20, twenty creators who get a solo operator most of the way on AI decisions. How we chose them. Facts above were checked against the creator's own pages on September 12, 2026.