| name | hiring-eval |
| description | Hiring evaluation from artifacts, repos, demos, resumes, or interview evidence. Use for "would you hire this engineer?", "top notch?", senior/staff calibration, HFT/systems candidate assessment, and what evidence would change the decision. NOT for production adoption of a project (use cto-eval) or business/demo ROI (use ceo-eval). |
| when_to_use | hiring eval, candidate evaluation, would you hire, top notch engineer, senior engineer, staff engineer, HFT role, systems engineer interview, portfolio review, repo as hiring signal, artifact evaluation |
| user-invocable | true |
Hiring Eval
Evaluate the person, not the product. The output is a hiring recommendation
grounded in evidence, with explicit uncertainty and a concrete bar for changing
the decision.
Dispatch
- Artifact only: evaluate signal quality and missing evidence.
- Artifact + repo: inspect the repo, run available checks, and sample core code.
- Interview notes: separate observed behavior from interviewer inference.
- HFT / low-latency / systems role: apply the HFT addendum below.
If the user asks for a "very critical" read, do not soften the conclusion.
Still distinguish "not proven" from "bad".
Evidence Order
Use the strongest available evidence first:
- Working code and tests you can run.
- Core implementation files and failure-path tests.
- Bench harnesses, raw data, and methodology.
- Architecture docs, bug logs, design notes, and runbooks.
- Demo videos, terminal casts, README claims, resumes, and prose.
Never treat a polished demo, README, or benchmark table as proof by itself.
Required Checks
For a repo-backed evaluation:
git status --short
make test
make build
make lint
Adapt commands to the project if no Makefile exists. If a command fails because
of local tooling, say exactly why, make the smallest reversible environment
adjustment needed, and retry when reasonable.
Then inspect:
- repo shape, commit recency, and ownership burden
- test count and ignored/skipped test categories
- core domain code, not just CLI/UI wrappers
- benchmark harnesses and benchmark reports
- bug tracker / TODO / known gaps
- docs that distinguish aspiration from measured fact
Judgment Dimensions
Score mentally; do not print a numeric score unless asked.
- Technical depth: understands the domain constraints, not just libraries.
- Correctness rigor: tests invariants, failure modes, replay/recovery,
edge cases, concurrency, and invalid inputs.
- Performance rigor: uses credible baselines, p50/p99/p999/max where
relevant, avoids coordinated omission, explains hardware and methodology.
- Taste: simple boundaries, few moving parts, no gratuitous abstraction,
honest tradeoffs.
- Operational maturity: observability, health checks, deploy model,
recovery runbooks, limits, backpressure, maintenance story.
- Communication: claims are precise, caveated, and falsifiable.
- Self-criticism: known gaps are named plainly and prioritized.
- Team risk: bus factor, maintainability, ability to work in a team,
overfitting to personal tooling, tendency to grandstand.
HFT / Low-Latency Addendum
For HFT, require evidence beyond "fast":
- tail latency: p99, p999, max, jitter, and saturation behavior
- benchmark setup: CPU model, kernel, governor, isolation, IRQ affinity,
NIC/driver if applicable, warmup, sample size, raw data
- measurement integrity: no coordinated omission, realistic baselines,
tuned competitors, repeatability
- hot-path discipline: allocation, syscalls, locks, atomics, cache layout,
false sharing, backpressure
- failure semantics: packet loss, gaps, replay, duplicate messages,
partial writes, crash recovery, clock issues, restart behavior
- exchange-domain correctness: price-time priority, order lifecycle,
risk checks, position accounting, liquidation/funding, audit trail,
exactly-once/at-most-once boundaries
- production constraints: kill switches, observability, deterministic replay,
deploy/rollback, incident handling
An HFT "top notch" hire must show both mechanical sympathy and paranoia about
correctness. Low p50 without tail/failure evidence is not enough.
Verdicts
Use one of these:
- HIRE STRONG — clear evidence of exceptional engineering for the target
role; remaining gaps are normal onboarding/domain gaps.
- HIRE — strong evidence; some risks, but likely worth hiring.
- ONSITE / DEEP INTERVIEW — promising, but not enough evidence for a hire
without targeted probing.
- NO HIRE FOR THIS ROLE — may be good, but evidence misses role-critical
requirements.
- NO HIRE — evidence shows poor judgment, weak fundamentals, or high team
risk.
Default to ONSITE / DEEP INTERVIEW when artifacts are impressive but the
role-critical proof is incomplete.
Output Template
Keep the answer direct:
**Verdict:** <verdict>
<1-3 sentence bottom line.>
Why:
- <evidence-backed point>
- <evidence-backed point>
Concerns:
- <role-critical gap or risk>
- <role-critical gap or risk>
What would change my mind:
1. <specific evidence>
2. <specific evidence>
3. <specific evidence>
If the user asked for a CTO-style hiring view, phrase the bar in business terms:
"Would I trust this person near the matching/risk path?" and "What interview
must de-risk before offer?"
Interview Probes
When useful, include 3-6 probes:
- Ask them to explain the hardest bug or wrong assumption in the project.
- Ask for a raw benchmark run and have them defend methodology.
- Pick a failure scenario and make them walk through exact state transitions.
- Ask where they would delete code.
- Ask what they would not ship to production yet.
- Ask them to compare against a serious alternative and name where theirs loses.
Rules
- NEVER infer "top notch" from polish alone.
- NEVER punish ambition by itself; punish unsupported claims.
- NEVER conflate product readiness with candidate ability.
- NEVER ignore maintenance burden or team-fit risk.
- ALWAYS say what evidence would change the verdict.
- ALWAYS separate observed facts from inference.