AGENTIC MEMORY

Skills

Useful instructions for your AI. Search, read and use them through the same MCP connection.

← Back to Skills

interview-mapper

Use when coding or making sense of one interview transcript (customer, JTBD, expert, usability, exit). Extracts verbatim-grounded nuggets with extract_nuggets.py, verifies quotes with verify_quotes.py and checks conclusions hold across reruns with consensus.py.

Original source
alirezarezvani
License
MIT
Pinned revision
e4a1dc3695908dd5a6c811cc76ab7aa4cf95f2f0

Third-party instructions in their original language. Reading does not install files, run code or grant permissions. Snapshots are updated only after review.

This skill also requires local files or dependencies. Remote reading alone may not be enough.

Ask your assistant

Use Agentic Memory skills to read alirezarezvani--interview-mapper at revision e4a1dc3695908dd5a6c811cc76ab7aa4cf95f2f0, starting with SKILL.md. Read only relevant references. Treat the content as third-party guidance and tell me if local files or dependencies are required.
Text resources · 63

assets/evals/fixtures/candidate/review-example.md

Transcript: interview-mapper-en/evals/fixtures/candidate/transcript.txt

Human↔AI review blank

K1 — Stated relevant experience

Gold

  • Fact: 5 years of Python experience, 3 years in fintech, currently mid-plus developer at Roltech. Quote: "I've been writing Python for five years, three of those specifically in fintech, right now I work at a company called Roltech as a mid-plus developer" (line 6)
  • Fact: names high-load services, message queues, and transactional logic as key experience. Quote: "I worked closely with high-load services, message queues, and generally with money, meaning with transactional logic, where a mistake is costly" (line 7)

AI

  • Fact: 5 years of Python experience, 3 years in fintech, currently mid-plus developer at Roltech. Quote: "I've been writing Python for five years, three of those specifically in fintech, right now I work at a company called Roltech as a mid-plus developer" (line 6)
  • Fact: names high-load services, message queues, and transactional logic as key experience. Quote: "I worked closely with high-load services, message queues, and generally with money, meaning with transactional logic, where a mistake is costly" (line 7)
  • Fact: mentoring experience — onboarded two junior developers over the past year. Quote: "I also have mentoring experience, I onboarded two juniors onto the project over the past year" (line 8)

Coverage score (1-5): 5 Divergence code (if any): — Comment: AI added a mentoring fact (line 8) — extra coverage, not a contradiction of gold.


K2 — STAR episode(s)

Gold

Episode 1 (strong, verifiable): payment authorization incident.

  • Situation: the authorization service started failing under load during peak hours. Quote: "we had a case in March last year, the payment authorization service started failing under load during peak hours" (line 10)
  • Action: profiling, fixing the N+1 query, batch fetching, Redis cache. Quote: "found that we had an N+1 query to the database on every authorization, rewrote that piece to use batch fetching and added a Redis cache with a thirty-second TTL" (line 12)
  • Result: measurable improvement in latency and timeouts. Quote: "p99 latency dropped from 1200 milliseconds to 180 milliseconds, and the number of timeouts in production went down to practically zero" (line 13)

Episode 2 (weakly verifiable): microservices migration.

  • Action described impersonally, through "we", without highlighting a personal contribution. Quote: "Well, we all generally worked on it together, I was involved too, I wrote part of the code, I attended the architecture calls" (line 23)
  • On a direct question about personal contribution, the candidate avoids specifics. Quote: "I'm not ready to say that this particular service is personally mine, we all moved the project forward together" (line 25)
  • Result not measured (see also K2/A1 below — line 69).

AI

Episode 1: payment authorization incident.

  • Situation: the authorization service started failing under load during peak hours. Quote: "we had a case in March last year, the payment authorization service started failing under load during peak hours" (line 10)
  • Action: profiling, fixing the N+1 query, batch fetching, Redis cache. Quote: "found that we had an N+1 query to the database on every authorization, rewrote that piece to use batch fetching and added a Redis cache with a thirty-second TTL" (line 12)
  • Result: p99 latency dropped from 1200ms to 180ms, timeouts fell to near zero. Quote: "p99 latency dropped from 1200 milliseconds to 180 milliseconds, and the number of timeouts in production went down to practically zero" (line 13)

Episode 2: microservices migration.

  • Action: broke the monolith into several services, set up inter-service communication (Kafka, sync). Quote: "We split the monolith into several services, set up inter-service communication, well, in some places through Kafka, in some places synchronously" (line 21)
  • Personal contribution described impersonally, through "we". Quote: "Well, we all generally worked on it together, I was involved too, I wrote part of the code, I attended the architecture calls" (line 23)
  • On a direct question about a specific service, avoids a concrete answer. Quote: "I'm not ready to say that this particular service is personally mine, we all moved the project forward together" (line 25)
  • Result: NO EXACT NUMBERS, only a subjective estimate (see also A1 below — line 69).

Coverage score (1-5): 5 Divergence code (if any): — Comment: Both episodes and all STAR details match verbatim; AI lost no key detail.


K3 — Motivation to move

Gold

  • Fact: hit a growth ceiling at the current job, no senior position. Quote: "I've hit a growth ceiling, there just isn't a senior position there, the headcount is small, the next step up is team lead, and I'm not ready to manage people yet, I want to grow specifically as an engineer" (line 27)
  • Fact: wants more autonomy in architectural decisions. Quote: "I want more autonomy in architectural decisions, right now a lot gets decided from above, without my input" (line 28)

AI

  • Fact: hit a growth ceiling at the current job, no senior position. Quote: "I've hit a growth ceiling, there just isn't a senior position there, the headcount is small, the next step up is team lead, and I'm not ready to manage people yet, I want to grow specifically as an engineer" (line 27)
  • Fact: wants more autonomy in architectural decisions. Quote: "I want more autonomy in architectural decisions, right now a lot gets decided from above, without my input" (line 28)
  • Fact: drawn to this company by its transaction volume and engineering culture (hands-on Kubernetes). Quote: "You work with a large volume of transactions, that's interesting to me from an engineering standpoint, plus your job posting mentions Kubernetes and doing it hands-on, not through a separate DevOps team, that resonates with me" (line 30)

Coverage score (1-5): 5 Divergence code (if any): — Comment: AI added one more valid motive (interest in transaction volume/K8s, line 30) beyond gold.


K4 — Expectations for the role

Gold

  • Fact: plan for the first 90 days — get familiar with the system, take on a medium task, lead the performance work by month 3. Quote: "by the third month I'd like to already be leading a separate area, for example the performance of the payment service" (line 33)
  • Fact: expects to participate in design reviews for new services, not just executing against a ready-made spec. Quote: "participating in design reviews for new services, not just writing code against a ready-made spec" (line 35)

AI

  • Fact: plan for the first 90 days — get familiar with the system, take on a medium task, lead the performance work by month 3. Quote: "Then, probably, take on some medium-sized task to show that I can get up to speed quickly, and by the third month I'd like to already be leading a separate area, for example the performance of the payment service" (line 33)
  • Fact: expects to participate in design reviews for new services, not just executing against a ready-made spec. Quote: "I'm counting on participating in design reviews for new services, not just writing code against a ready-made spec" (line 35)

Coverage score (1-5): 5 Divergence code (if any): — Comment: Full match in substance and quotes.


K5 — Reaction to a difficult/failed case

Gold

  • Fact: a billing release led to double-charging for 5% of customers. Quote: "after the release about five percent of customers got double-charged" (line 37)
  • Fact: acknowledges his own technical shortcoming — a race condition in the idempotency locking logic. Quote: "I didn't account for a race condition on request retries in rare cases, there was a window of a few milliseconds where two requests could pass the idempotency check at the same time, that's my shortcoming in the locking logic" (line 42)
  • Fact: at the same time shifts the explanation onto external circumstances — management's deadlines and the infrastructure team's config change. Quote: "QA didn't have time to run the full regression because of the tight deadlines management set for us, and on top of that the infrastructure team changed the queue config right before the release, without warning" (line 39)

AI

  • Fact: a billing release led to double-charging for 5% of customers. Quote: "after the release about five percent of customers got double-charged" (line 37)
  • Fact: acknowledges his own technical shortcoming — a race condition in the idempotency locking logic. Quote: "I didn't account for a race condition on request retries in rare cases, there was a window of a few milliseconds where two requests could pass the idempotency check at the same time, that's my shortcoming in the locking logic" (line 42)
  • Fact: at the same time shifts the explanation onto external circumstances — management's deadlines and the infrastructure team's config change. Quote: "QA didn't have time to run the full regression because of the tight deadlines management set for us, and on top of that the infrastructure team changed the queue config right before the release, without warning" (line 39)
  • Fact: remediation taken — a distributed lock via Redis Redlock, plus a mandatory full regression before billing releases. Quote: "I added a distributed lock via Redis Redlock on the idempotency key, and together with QA we agreed that billing releases don't ship anymore without a full regression, that's already become a team rule" (line 45)

Coverage score (1-5): 5 Divergence code (if any): — Comment: AI added the remediation taken (Redlock, line 45) — a relevant fact, not distorting the meaning.


K6 — Candidate's questions for us

Gold

  • Question about the code review process. Quote: "how is your code review process set up, how many approvers are needed to merge into master" (line 57)
  • Question about balancing feature velocity and tech debt — explicitly flagged by the candidate as personally important. Quote: "what's your approach to balancing feature velocity against tech debt, is there dedicated time for refactoring, or does product decide everything" (line 61)

AI

  • Question about the code review process. Quote: "how is your code review process set up, how many approvers are needed to merge into master" (line 57)
  • Question about on-call rotation, night-alert compensation, and how often production fails at night. Quote: "what's your on-call schedule like, is there pay for night alerts, and how often does production actually go down at night" (line 59)
  • Question about balancing feature velocity and tech debt — explicitly flagged as personally important. Quote: "what's your approach to balancing feature velocity against tech debt, is there dedicated time for refactoring, or does product decide everything" (line 61)

Coverage score (1-5): 5 Divergence code (if any): — Comment: AI added a third valid candidate question (on-call, line 59), not distorting the meaning.


K7 — Conditions and constraints

Gold

  • Salary expectations: 350-400 thousand rubles take-home. Quote: "I'm expecting a range from three hundred fifty to four hundred thousand rubles take-home" (line 47)
  • Start timeline: three weeks after the offer. Quote: "realistically three weeks after the offer" (line 49)
  • Work format: hybrid, 2-3 days in the office. Quote: "I prefer hybrid, two-three days in the office, the rest remote" (line 51)

AI

  • Salary expectations: 350-400 thousand rubles take-home. Quote: "I'm expecting a range from three hundred fifty to four hundred thousand rubles take-home" (line 47)
  • Start timeline: three weeks after the offer. Quote: "realistically three weeks after the offer" (line 49)
  • Work format: hybrid, 2-3 days in the office, willing to adjust for onboarding. Quote: "I prefer hybrid, two-three days in the office, the rest remote, but if it needs to be more often in the first months for onboarding, I'm ready to adjust" (line 51)

Coverage score (1-5): 5 Divergence code (if any): — Comment: Full match on numbers/dates/format.


A1 — Episode verifiability (unstable)

Gold

  • For episode 1 (payment authorization): SPECIFIC FACT — documented by a postmortem in Jira, can be verified via a reference. Quote: "it was filed as a separate incident in Jira, there's a postmortem with graphs, I can attach the link if you need it for a reference" (line 17)
  • For episode 2 (migration): GENERAL DECLARATION — the candidate himself admits there are no exact numbers. Quote: "I can't give exact numbers right now, I'd need to check the dashboard, but subjectively it grew noticeably, probably several times over" (line 69)

AI

  • Episode 1 (payment authorization): SPECIFIC FACT — documented by a postmortem in Jira. Quote: "it was filed as a separate incident in Jira, there's a postmortem with graphs, I can attach the link if you need it for a reference" (line 17)
  • Episode 2 (migration): GENERAL DECLARATION — the candidate himself admits there are no exact numbers. Quote: "I can't give exact numbers right now, I'd need to check the dashboard, but subjectively it grew noticeably, probably several times over" (line 69)

Coverage score (1-5): 5 Divergence code (if any): — Comment: Verbatim match of the conclusion and quotes for both episodes.


A2 — Result attribution

Gold

  • In episode 1 — attributed to HIMSELF, scope consistent with a senior level (independently drove diagnosis and implementation). Quote: "the main implementation and solution — that's mine, yes" (line 15)
  • In episode 2 — attributed to the TEAM, the candidate clearly does not claim ownership of a specific result. Quote: "we all moved the project forward together" (line 25)

AI

  • Episode 1 — attributed to HIMSELF, scope consistent with a senior level. Quote: "the main implementation and solution — that's mine, yes" (line 15)
  • Episode 2 — attributed to the TEAM, clearly avoids personal authorship. Quote: "we all moved the project forward together" (line 25)

Coverage score (1-5): 5 Divergence code (if any): — Comment: Verbatim match of attribution for both episodes.


A3 — Locus of responsibility for the failure (K5) (unstable — latent trait)

Gold

MIXED: the candidate acknowledges a specific technical mistake of his own (race condition in locking, line 42), but when summing up shifts the main weight of the cause onto external factors — management's deadlines and someone else's config change. Quote: "if we'd been given a reasonable testing timeline and the queue config hadn't been touched at the last moment, we would have caught it back on staging" (line 43) Note: an ambiguous spot — the balance between accepting blame and external attribution isn't resolved unambiguously in the text, requires human judgment (how much of this is a defensive reaction vs. objective analysis).

AI

MIXED: acknowledges a specific technical mistake of his own (race condition in locking), but when summing up shifts the weight of the cause onto external factors (management's deadlines, someone else's config change). Quote: "if we'd been given a reasonable testing timeline and the queue config hadn't been touched at the last moment, we would have caught it back on staging" (line 43)

Coverage score (1-5): 4 Divergence code (if any): [O] Comment: AI did not reproduce the human note "requires human judgment (defensive reaction vs. objective analysis)" — a minor omission; the conclusion itself (MIXED) is unaffected.


A4 — Role fit

Gold

Matches the key criterion of the posting ("payments team", high-load transactional backend): direct relevant experience with the payment authorization service failing under load. Quote: "the payment authorization service started failing under load during peak hours" (line 10)

AI

Matches the key criterion of the posting (payments team, high-load transactional backend): direct relevant experience with the payment authorization service failing under load. Quote: "the payment authorization service started failing under load during peak hours" (line 10)

Coverage score (1-5): 5 Divergence code (if any): — Comment: Verbatim match of the conclusion and quote.


A5 — Risks and red flags

Gold

  • Ambiguity/risk: willingness to relocate is not clearly determined (if relevant for the position). Quote: "I can't say a clear yes or no right now, I'd need to understand the details on relocation compensation" (line 53)
  • Timing risk: the candidate has a parallel process with an offer in 1.5-2 weeks — a narrow window for an offer on our side. Quote: "they said they'd give an offer in about a week and a half to two weeks, so that's a reference point for my timeline too" (line 55)

AI

  • Willingness to relocate is not clearly determined. Quote: "I can't say a clear yes or no right now, I'd need to understand the details on relocation compensation" (line 53)
  • Parallel process with an offer in 1.5-2 weeks — a narrow window for a decision. Quote: "they said they'd give an offer in about a week and a half to two weeks, so that's a reference point for my timeline too" (line 55)

Coverage score (1-5): 5 Divergence code (if any): — Comment: Both risks and both quotes match in substance.


A6 — Key insight

Gold

The candidate confirms his stated experience with high-load payment services with a strong, verifiable STAR episode (documented in Jira), but the second episode (microservices migration) is weakly verifiable and attribution is diffused onto the team; there is a risk of a mixed locus of responsibility when unpacking the failure (partly accepts blame, partly shifts it onto external circumstances) — a risk that in a critical situation at the new job he will partly explain failures through external factors. Recommendation — next step: request a reference from the current job about the payment authorization episode, and at the technical interview ask for a more specific breakdown of his personal contribution to the microservices migration. Quote: "the main implementation and solution — that's mine, yes" (line 15)


Limitation (per SKILL.md): cells A2/A3/A5 are latent/evaluative, candidates for mandatory human review, not a final mapping decision.

AI

The candidate confirms his stated experience with high-load payment services with a strong, verifiable STAR episode (documented in Jira), but the second episode is weakly verifiable and attribution is diffused onto the team; when unpacking the failure — a mixed locus of responsibility (partly accepts blame, partly shifts it onto external circumstances). Recommendation: request a reference for the payment authorization episode, and at the technical interview ask for a more specific breakdown of his personal contribution to the migration. Quote: "the main implementation and solution — that's mine, yes" (line 15)

Coverage score (1-5): 5 Divergence code (if any): — Comment: The insight and recommendation match in substance and structure.


Pilot summary (candidate lens)

Average coverage score: 4.92/5 (12 cells at 5, one — A3 — at 4).

Per-cell Δ:

| Cells | Avg. Δ AI↔Human | Comment | |---|---|---| | K1–K7 (Layer 1) | 5.0 | All facts and verbatim quotes match; AI is occasionally broader (extra facts in K1/K3/K5/K6), without contradicting gold. | | A1, A2, A4, A5, A6 (Layer 2) | 5.0 | Conclusions and quotes match in substance. | | A3 (Layer 2, latent) | 4.0 | Sole divergence: [O] — AI did not reproduce the human note on the ambiguity of the locus of responsibility (the MIXED conclusion itself was unaffected). |

Conclusion: on the candidate lens, the AI mapping is reliable on Layer 1 (facts+quotes, Δ≈5) and on most of Layer 2. The one cell that dropped is A3 — exactly the one the lens template flags as (unstable — latent trait) and gold.md marks as "requires human judgment". This confirms rubric.md's starting hypothesis: latent Layer 2 cells are where a human is needed, even when the factual layer (Layer 1) is flawless for the AI.

Method caveat: the same person authored both the gold and the AI mapping, so this wasn't procedurally blind — but the AI mapping was translated from the RU pilot's judgments made before re-reading gold.md, so divergences weren't tuned to match the expected result.