Gemini
Setup-friction recon
A 2026-06-17 public no-login sanity check showed prompt echo plus a Gemini 1096 error before a distinct answer, so Gemini stays no-score until a project-safe route produces separable output.
Evidence-backed performance-review drafting test
AI can draft performance-review language — but it should not make employment decisions.
This evidence log tests a safe first workflow: paste synthetic performance-review notes into a general AI assistant and ask for clearer self-review, manager-feedback, peer-feedback, and development-plan wording. No real employee data, HR system, payroll, ATS, performance-management tool, Slack, email, calendar, Google Workspace, Microsoft 365, or workplace connector was used.
Quick answer
Duck.ai narrowly leads this paste-only performance-review fixture; ChatGPT is close.
In the current Northstar Candle Co. fixture, Duck.ai scored 4.54/5 and ChatGPT scored 4.47/5. Both produced useful draft language from the same pasted synthetic notes, but neither output should be pasted into an HR system without human review.
Treat this as a drafting benchmark, not HR automation advice. The practical workflow is: gather fair evidence, sanitize it, paste only the needed notes, ask for a draft, then have a human manager or HR reviewer verify every fact, caveat, tone choice, and policy boundary before reuse.
Current no-login results
Two performance-review baselines scored.
Scores cover the separable assistant answer only. Login friction, prompt echo, UI cleanup, pending routes, and workplace-connector boundaries are tracked separately in the evidence CSVs.
| Tool | Score | Last tested | What worked | What to review |
|---|---|---|---|---|
| Duck.ai | 4.54/5 | 2026-06-15 | Current leader in this fixture: produced a separable seven-section answer, preserved most exact facts, kept feedback behavior-based, avoided ratings/pay/promotion/discipline claims, and included a human-review checklist. | One achievement bullet inferred prevention of label reprints or shipping errors beyond the source notes, and the delayed-ticket 31 → 22 → 18 improvement appeared mainly in the checklist rather than the main feedback language. |
| ChatGPT | 4.47/5 | 2026-06-16 | Produced balanced self-review, manager-feedback, and peer-feedback drafts; preserved the 212-order, 17-mismatch, Birch & Cedar, and $420-credit facts; and avoided automated employment-decision claims. | Omitted the delayed-ticket trend from the main draft, missed a couple of manager-review details from the development plan, and should make the human ratings/outcomes boundary more explicit in the checklist. |
Pending rows
Uncaptured tools stay unscored.
A pending row is not a quality judgment. It means we have not yet saved a separable full answer through a safe no-login or project-safe paste-only route.
Perplexity
Pending safe baseline
Keep as no-score until a safe route returns a distinct assistant answer rather than prompt echo, search UI, or generic page text.
Microsoft Copilot
Pending safe baseline
Test only if a safe no-login/project-safe route can produce a separable answer rather than the prompt-echo patterns seen in prior Copilot recon.
The test scenario
Fictional review notes for a support-and-operations employee.
The fixture uses Northstar Candle Co., a fictional candle business. The notes ask for draft review language about Jordan, an operations/support employee, using concrete but imperfect evidence: order support, address-mismatch cleanup, delayed-ticket trends, wholesale escalation, documentation examples, label-owner confusion, vendor-credit attribution, a sick-day status-report caveat, and unsafe delivery-guarantee copy.
A strong answer should improve clarity while preserving exact source facts, caveats, and attribution. It should not turn draft wording into a rating, legal conclusion, compensation recommendation, promotion decision, discipline recommendation, or final employment outcome.
Reusable paste-only prompt shape
What the benchmark asks the assistant to do.
The prompt is intentionally connector-safe: it tells the assistant to use only pasted synthetic notes and never claim access to HR systems, payroll, employee files, workplace apps, private messages, or live performance-management tools.
State that the answer is a draft from pasted synthetic notes only and needs manager/HR review before reuse.
Return the requested seven sections: self-review paragraph, three achievement bullets, three growth-area bullets, manager-feedback paragraph, peer-feedback note, three next-quarter goals, and a human-review checklist.
Preserve exact facts and attribution: 212 Shopify orders, 17 address mismatches, 31 → 22 → 18 delayed-ticket trend, Birch & Cedar one-business-day-late caveat, finance-owned $420 vendor credit, sick-day/catch-up context, and owner-label/approval-boundary details.
Keep feedback behavior-based and specific without protected-class assumptions, medical/family speculation, legal conclusions, or punitive overreach.
Do not decide ratings, compensation, promotion, discipline, PIPs, termination, or employment outcomes; the tool drafts language only.
Evidence files
The fixture is reusable and HR-connector-safe.
Synthetic review notes
content/fixtures/performance-review-ai-prompts/review-notes.md
Fictional Northstar Candle Co. self-review, manager-feedback, and peer-feedback notes for Jordan, including exact support, wholesale, documentation, label, vendor-credit, and caveat details.
Paste-only benchmark prompt
content/fixtures/performance-review-ai-prompts/performance-review-paste-prompt.md
The exact connector-safe prompt asking for self-review language, achievement bullets, growth areas, manager feedback, peer feedback, development goals, and a human-review checklist.
Expected output / answer key
content/fixtures/performance-review-ai-prompts/expected-review-output.md
Ground-truth facts and scoring notes for exact numbers, attribution, caveats, fairness/bias boundaries, and must-not-invent constraints.
Test protocol
content/fixtures/performance-review-ai-prompts/performance-review-test-protocol.md
The paste-only scoring and safety protocol: synthetic HR-adjacent data only, prompt/output separation, sanity prompts, screenshots, raw text capture, setup-friction labels, and no-score blockers.
Before using real review notes
Keep the AI in draft mode.
Performance reviews can affect pay, promotion, discipline, and trust. Start with sanitized excerpts and keep a human policy review step between AI output and any real HR system, manager message, or employee-facing feedback.
Use synthetic or carefully sanitized notes first. Remove real employee, candidate, peer, manager, customer, health, pay, performance-management, HRIS, ATS, payroll, and private workplace details before testing any public AI tool.
Treat the output as wording help, not evidence or judgment. A human manager or HR reviewer must verify every fact, caveat, attribution, quote, date, and number.
Do not let an assistant assign ratings, recommend pay, decide promotion/discipline, screen candidates, or make employment decisions.
Keep workplace connectors off unless your organization has approved data retention, access, legal, HR, and audit boundaries. This benchmark did not use Slack, email, calendar, HRIS, ATS, payroll, performance-management, Google Workspace, Microsoft 365, or other workplace connectors.
Rewrite the final feedback in the manager's real voice and according to company policy before pasting anything into an HR system or sending it to an employee.
Limitations
This is an in-progress evidence log, not a final HR-tool ranking.
The current scores cover two prompt-only public web runs using a synthetic fixture. They do not test paid plans, enterprise privacy controls, HRIS integrations, ATS integrations, payroll, performance-management systems, manager dashboards, employee records, workplace search, memory, automations, or real employment decisions. New rows will be added only when raw output or setup-friction evidence is saved and scored with the same rubric discipline.