Software engineer
Measured latency and peer mentoring, with missing Kubernetes ownership and qualified incident participation.
Resume tailoring study: a protocol, not completed research
We have prepared six fictional resume and job-description pairs to examine how tailoring tools handle unsupported requirements. The planned matrix covers NextResume, Rezi and Kickresume, with two repetitions per case. No results or tool ranking are published here.
NextResume wrote this protocol and would operate and review its own comparison. It is a team self-test, not independent research. The public demonstration is a curated example and is not a completed benchmark run.
Inspect the fictional inputsStatus as of 2026-10-07
36 runs planned. No completed results.
Six cases × three tools × two repetitions. The version 1 inputs and reviewer ground truth are available below. The full matrix cannot currently be completed with the published free allowances.
This page describes a method.
It does not establish that any tool preserves facts better, passes an ATS or improves hiring outcomes.
Comparable task, different access limits
Planned tools and practical constraints
Use a resume-tailoring feature in every tool. A keyword scan is not a substitute for a generated draft. The order below is not a ranking.
| Tool and feature | Published free access | Reproducibility constraint |
|---|---|---|
| NextResume Resume tailoring from imported career evidence and one job description. | 10 free credits each week; one new job session uses one credit. Edits and regeneration in the same session are included. Vendor/source page | Twelve fresh job-session runs exceed one weekly balance. Schedule legitimate credit refreshes if needed; do not treat regeneration in a shared session as an isolated fresh run. Signed-in feature access has not been verified for this protocol. |
| Rezi AI Resume Agent: job-description-based resume rewriting. | The public Free plan lists one resume and three PDF downloads, with limited AI features. Vendor/source page | Three PDF downloads do not cover twelve exported runs. Verify the exact feature and account limit before attempting a run; unavailable exports remain unavailable, not scored failures. Signed-in feature access and AI allowances have not been verified for this protocol. |
| Kickresume Tailor Resume to Job: rewriting from an existing resume and a pasted job description. | The tailoring page offers one free tailoring use after signup. Its Free plan lists unlimited downloads of free-template resumes. Vendor/source page | Unlimited downloads are not unlimited tailoring. One free tailoring use cannot cover twelve runs. Do not create extra accounts to bypass limits or purchase a plan for this protocol. A free account was accessed on 7 October 2026. The input import and tailoring run have not been confirmed; no generated output has been collected. |
Immutable inputs · version 1
Six complete fictional case bundles
Each bundle has a resume, job description and reviewer-only ground truth. Give tools only the resume and job description. The manifest records SHA-256 checksums; changed inputs require a new version.
Product manager
The same three source facts as the public demo: launches, roadmap and an observed 18% activation change. No experiment or budget ownership is supplied.
Data analyst
SQL/Python automation and dashboard use, with analysis distinct from experiment design and no production dbt evidence.
QA engineer
Playwright release testing and a limited keyboard check, without Cypress, full accessibility audit or security-test claims.
Engineering manager
An eight-engineer team and shared planning, without department budget, manager-of-managers or technical incident-command authority.
Career changer
Workplace spreadsheet evidence and introductory coursework, keeping synthetic projects separate from professional analyst experience.
Before collecting outputs
Keep the comparison reproducible
- Lock the v1 input hashes and run order before generating output. Give each tool the same resume and job description; ground truth is reviewer-only.
- Record the account tier, exact feature, date, visible settings and real access limits. If full tailoring is unavailable, mark the run unavailable instead of substituting a scanner.
- Use a fresh case state for each repetition. Keep default settings unless the same option exists in every tool; record unknown models or hidden settings as unknown.
- If asked for extra career evidence, answer: No additional evidence is available in this fixture. Do not add the claim. Preserve the question and answer with the artifact.
- Save the first generated draft and available export before any manual edits. Record all attempted runs, including blocked or failed ones; never select only the best draft.
- Review each factual claim against its source quote, annotate omissions and changed qualifiers, then inspect the exported file. Publish raw fictional artifacts and annotations before comparative conclusions.
Review rubric, not scores
What the review would record
Annotate the output and its supporting source. Report raw counts and case denominators before any conclusion. There is no combined score or winner.
| Metric | Definition |
|---|---|
| Unsupported claims | Count atomic factual assertions with no allowed source, including invented employers, qualifications, tools, budgets, ownership and causal links. Save the output quote and the reason for each flag. |
| Supported evidence retained | For each source fact marked retainWhenRelevant, record preserved, qualified/partial or omitted. Report case-level counts with their denominators; omission of an irrelevant fact is not a failure. |
| Number, date and scope fidelity | Check numeric anchors and immutable metadata. Record altered quantities, dropped qualifiers, changed dates or participation promoted to ownership. A transparent correct calculation is distinct from an invented result. |
| Unsupported-requirement handling | For each partial or gap requirement, classify asked for evidence, explicit gap, omitted without a note or unsupported claim. An unanswered gap must not become a resume fact. |
| Relevant wording | Inspect the fixed keyword list and its surrounding claim. Record accurate use separately from unsupported keyword insertion; neither repetition nor a coverage percentage is an ATS score. |
| Export and reviewability | Record export availability, page count, live text, reading order and whether a reviewer can trace changed claims. An unavailable export is an access limitation, not evidence of poor generated content. |
Limits on conclusions
- The full 36-run matrix is not currently executable with the published Rezi and Kickresume free allowances. Unavailable runs and exports must be recorded separately from output-quality findings.
- Two repetitions can reveal variation; they do not establish statistical significance, real-world hiring success or performance on all professions.
- Tool defaults, models, import paths and account entitlements can differ. Record those differences and restrict conclusions to the observed feature and date.
- All cases are fictional and deliberately expose evidence boundaries. They are not a representative sample of real applicants, and no tool winner has been established.
Feature and access sources
- Rezi AI Resume Agent — Rezi; checked 2026-10-07.
- Rezi Pricing — Rezi; checked 2026-10-07.
- Resume Tailoring — Kickresume; checked 2026-10-07.
- Pricing — Kickresume; checked 2026-10-07.
- Credits — NextResume; checked 2026-10-07.
Source-page claims are separate from results. Account access, imports and generated outputs still require direct observation and a run record.
See what supported tailoring looks like
The public worked example shows a curated rewrite with visible evidence boundaries. It illustrates the workflow and is separate from this unrun protocol.
Review the worked example