The signals were there, but she was too busy to hear them.
Last spring, Priya hired the most impressive candidate of her career. He was articulate and relaxed, and he had an answer for everything. After 6 months, the project that he led was 3 months late. His team had quietly stopped bringing him their problems.
When she reread her interview notes, it was all there. Every story was told with "we" and never with "I". The stories didn't have any numbers in them. Every hard question had been answered with a bigger word. Priya was a good interviewer. She had missed those signals because, during the interview, she was listening, writing notes, watching the clock and thinking of her next question, all at the same time.
Now she had to replace him. She had 12 candidates to interview in 2 weeks, on top of running a platform team of 40.
"I wanted to do the judging myself, with a second pair of ears that never got tired."Priya
Chapter 2 · Teaching it what "great" looks likeThe setup took 40 minutes on a Sunday evening.
Priya created a Scenario called Staff Engineer · Interviews and filled its folder with everything that lived in her head and on her drives:
- The job description, and the scorecard. The scorecard has 5 competencies (technical judgment, ownership, communication, mentoring, delivery), and for each one it describes what levels 1 to 4 sound like.
- "What great looks like": 3 paragraphs about the best Staff Engineer who she had ever worked with.
- Green flags: talks in trade-offs, owns failures with an "I", measures outcomes, asks about the team. Red flags: says "we" and never "I", uses buzzwords without depth, blames others, never says "I don't know".
- The architecture overview of her platform, so that the suggestions would fit her real situation. She also added her question bank.
Then she wrote the most important line in START-HERE.md: "I'm the interviewer. Don't answer for the candidate. Tell me my best next question, and map what you heard to the scorecard with the exact quote." She insisted on one more line: "Use evidence only. Never assume anything from age, accent, gender or background." Finally, she changed one setting. Each answer could now read the whole hour of conversation instead of the last 5 minutes, so that her summary at the end of the interview would cover everything.
Chapter 3 · The interviewsCandidate 4 told a great story, but Candidate 9 told a true one.
Candidate 4 was polished. He described a migration to Kubernetes that "went really well overall". It was the kind of answer that had fooled her before. Priya pressed her keyboard shortcut, and half a second later her screen showed this:
- ONSUBJECT.AI · 0.5 sAsk: "What was the hardest call that you made in that migration, and what did it cost?" · So far he has said "we" 9 times and "I" once, without any numbers → ask about ownership.
She asked the question. After a long pause, he told a story about "alignment with stakeholders". The next suggestion was "If you had to do it again with half the budget, what would you cut first?" His answer to it told her everything that she needed to know.
Candidate 9, Amina, was the opposite. She was quiet, almost shy. On a tired Thursday afternoon, her calm delivery could easily have read as "low energy". But she talked about idempotency keys without being asked. She also described the night when she personally ordered a rollback at 2 a.m. She chose a 40-minute outage over 2 days of repairing data.
- ONSUBJECT.AI · 0.5 sGreen: owns a hard call and its cost → Ownership, level 4. Quote: "I made the call to roll back at 2 a.m." · Ask next: "Who else can run that runbook now?" (mentoring: not shown yet)
Priya hadn't written a single note in 10 minutes. She had been listening. For the first time in years, her ninth interview of the week was as sharp as her first.
Chapter 4 · After the callThe debrief took 15 minutes instead of 45.
When each interview ended, Priya typed one line into the box under the transcript: "Summarise the evidence against the scorecard, with quotes: strengths, concerns and what to verify in references." She had a fair, structured write-up 2 seconds later. It was built only from what had actually been said.
Then she cleared the transcript for the next candidate. Each interview was already saved in her Scenario, with one file per candidate. Each file had every line with its time and speaker, and every suggestion marked [OnSubject.AI] at the point where it happened. In the hiring debrief, the team discussed what the candidates had really said, so it didn't matter who remembered it best.
"For the first time, my best interview of the week and my last one were equally good."Priya, 8 months after hiring Amina
Interviewing fairly with AI. Tell candidates that the conversation is transcribed for your notes. In many places it's a legal requirement, and it's always good practice. OnSubject.AI suggests questions and organises the evidence. The decision is yours, and it should rest on job-related criteria that are applied equally to everyone. Write that into your Scenario, as Priya did.
An illustrative story: the names and companies are fictional. OnSubject.AI works the way the story describes.