Making the people who don't present well visible

What actually happens in the room

Any cross-team review covering a few dozen projects ends in the same argument: was the winner the best work, or the best pitch?

Nobody’s at fault. A judge sitting through thirty presentations in a day has finite attention, so whoever is clear, well-paced and quotable gets an edge. And there’s a subtler layer: presenting rewards talking about plans. There’s only one thing you actually finished, and infinite things you might do next — filling the last three slides with a roadmap simply looks better than saying “we got one scenario working”.

AI Achievement Interview (openagent.world) is aimed at exactly those two failure modes.

Three hard rules

The site puts them in the most prominent spot on the page:

• don’t grade eloquence
• don’t count plans as results
• score with one model, one ruler

The full version of the third: every project is scored with the same model, the same prompt and the same weights, with no manual adjustment this round.

“No manual adjustment this round” reads like a self-imposed limitation. It’s actually the only point the whole method can stand on — because manual adjustment is precisely the door through which subjective variance walks back in.

One interview, 1–2 minutes

Three steps:

01 voice confirmation    02 highlight extraction    03 unified scoring
name·dept·project        facts·artefacts·open       six-dim total·ranking

One question at a time. The AI decides from your answer whether to clarify, chase the event, chase the change or chase the boundary, rather than mechanically working through a question bank. This isn’t a fixed set of questions; it’s a method for talking about outcomes, taking the shortest path to three things:

• one real event
• one inspectable piece of evidence
• one question still open

The last one gets overlooked, and it’s the strongest authenticity signal there is: anyone who actually did the work can tell you what they haven’t cracked yet.

Also: facts, judgements and plans are recorded separately, and anything thin is explicitly marked “to be confirmed” — so vague statements can’t drift into the finished column.

Two routes, one ruler

On-site and remote run in parallel, but the standard doesn’t change with the route:

• On site — staff record from a desktop dashboard while ChatGPT Voice on a phone handles questions and adaptive follow-ups
• Remote, self-serve — start from the web, no admin-issued link and no ChatGPT client; if the realtime service falters it falls back to a standard six-question mode

Both end up in the same candidate queue, the same transcription pipeline and the same six-dimension scoring.

Recording, transcript, highlights, score and report all share one candidate-project identifier. That reads like an implementation detail, but anyone who has run a review knows that the moment one person’s recording gets attached to another person’s project, the credibility of the whole round is gone.

Letting people redo it lowers the weight on eloquence

You can re-record as many times as you like before submitting, at no cost. After a formal submission you can still redo the interview, up to 3 times total — the latest recording, transcript and score count, while every previous recording, realtime Q&A log and action record is kept.

This is the other face of “don’t grade eloquence”: with only one shot, nervous people lose. Once you can redo it, “spoke fluently” naturally stops carrying weight, and what’s left is what you actually did.

Keeping the earlier attempts serves a different goal — traceability. Scoring the latest one is a rule about which number counts, not a reason to erase the rest.

Where it fits

Organisations that need to compare dozens or hundreds of projects in a short window, and don’t want the outcome to hinge on who presents best.

Worth reading the preparation checklist on the site first: knowing what it will ask is what lets you make those 1–2 minutes concrete.