Real workload trials
No toy examples. Marketing prompts run against live ad accounts and real landing pages. Coding prompts run against production repositories. If the output needs a full rewrite, the prompt fails.
About PromptForge
A small team with unreasonable standards
PromptForge started in 2024 inside a marketing agency in Hamburg. We kept a shared doc of prompts that actually worked — the ones that survived contact with real client briefs, real deadlines and real code review. That doc grew to 400 entries, then colleagues asked for copies, then strangers did.
Today we're a fully remote team of five: two former agency operators, an ex-Google ML engineer, a technical writer, and a designer who thinks in Midjourney parameters. Between us we've run prompts against millions of tokens of production work — ad accounts, codebases, legal summaries, editorial calendars.
We sell prompt packs because we believe the gap between 'AI is amazing' and 'AI actually shipped this' is almost always prompt quality. Closing that gap is the whole company.
Methodology
Tested against real workloads, or it doesn't ship
Every prompt in every pack survives the same three-stage gauntlet. Roughly 40% of candidates fail. The ones that pass are the ones you buy.
Adversarial inputs
We feed each prompt vague briefs, missing context and contradictory instructions — because that's what real users paste. Prompts must ask clarifying questions or state assumptions instead of hallucinating confidence.
Blind cross-model scoring
Outputs are scored blind by the team across at least three frontier models. A prompt must beat the naive one-line version by a clear margin — otherwise you're paying for something you could have typed yourself.
Quality standards
What 'professionally engineered' means here
Versioned like software
Every prompt carries a version number and changelog. When a model update shifts behaviour, we re-test the whole pack.
Failure modes documented
Each prompt lists where it breaks: contexts it's bad for, inputs that degrade it, and the fix.
Structure over vibes
Prompts encode frameworks — positioning methods, review checklists, decision trees — not 'act as a rockstar' roleplay.
No filler
If a pack can be excellent at 85 prompts, we don't pad it to 200. Counts on the cards are the real counts.
The team
Five people, four time zones, one obsession
NK
Nadia Krüger
Founder · Prompt methodology
JM
Jonas Meireles
ML engineering · Eval systems
AO
Aiko Otanashi
Design · Midjourney pipelines
RB
Rory Byrne
Writing · Editorial packs
LS
Lena Stojanović
Operations · Customer success