Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Use when auditing a CVPR submission before the OpenReview deadline, covering abstract registration and profile requirements, the 8-page limit including figures and tables, the Compute Reporting Form, anonymity and external-link bans, dual-submission attestations, reviewer-duty enrollment, and desk-reject triage.
.claude/skills/brycewang-stanford-cvpr-submission/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-01 | ✗→✓ | ▲ Improved | 69% | 0% |
| case-02 | ✗→✓ | ▲ Improved | 15% | 0% |
| case-05 | ✗→✓ | ▲ Improved | 70% | 0% |
| case-06 | ✗→✓ | ▲ Improved | 43% | 0% |
| case-08 | ✗→✓ | ▲ Improved | 86% | 0% |
Run this audit in the two weeks before the CVPR paper deadline. All specifics below are the CVPR 2026 cycle as read on 2026-07-08 (paper deadline November 13, 2025); the next Program Chairs will rewrite details, so reopen the current Author Guidelines and Dates pages at cvpr.thecvf.com before acting on any number here.
CVPR staggers three hard cut-offs a week apart, and the first one is the one teams miss:
Every listed author — including the advisor added "later" and the intern who wrote one module — needed a complete, valid OpenReview profile by this date. An incomplete profile on any coauthor is desk-reject exposure for the paper.
author list, subject areas, conflicts, and the Compute Reporting Form all land here.
supplement; nothing in the main PDF can change after the paper deadline.
The cap is 8 pages including figures and tables, with extra pages permitted only for cited references. That "including figures" clause is what distinguishes CVPR budgeting from text-limited venues: a half-page teaser figure, two qualitative grids, and three result tables can consume four of your eight pages. Plan figures as page-budget line items, not decorations, and treat any style-file edit, margin nudge, or font substitution as a desk-reject trigger — over-length and mis-formatted papers are rejected without review. Build from the official author kit (github.com/cvpr-org/author-kit), current year's branch, unmodified.
New in the 2026 cycle: every submission carried a CRF. Only Section 1 (hardware) and Section 5 (verification) were mandatory — about five minutes of work — and a documented opt-out path existed via the comments field for proprietary or unrecorded compute. Fill the mandatory sections early; do not let a five-minute form become a deadline-night blocker, and do not fabricate GPU-hour numbers to look either frugal or mighty.
CVPR is double-blind with two venue-specific wrinkles. First, arXiv posting is not an anonymity violation — you may have a preprint up, and reviewers are told not to hunt for it. Second, external links are banned when they expand content or subvert review: an anonymous-looking project page, a Google Drive of extra results, or a demo URL can breach anonymity, the media embargo, or the length limit all at once. Everything reviewers need must live in the PDF and the uploaded supplement.
bash# Mechanical sweep on the final build, ten minutes before upload pdftotext paper.pdf - | grep -nEi 'github\.com|drive\.google|project page|our (prior|previous) work' | head pdfinfo paper.pdf | grep -Ei 'author|creator|producer' # identity in metadata grep -rn 'vspace\|\\addtolength\|geometry' main.tex # style-tampering tells pdftotext paper.pdf - | awk 'END{print NR" lines extracted"}' # sanity: text present, not raster
By submitting you attest that no substantially similar paper is under review at another conference or workshop during the defined review period (Nov 13, 2025 – Feb 20, 2026 in the 2026 cycle). Two verified carve-outs: prior papers of four pages or fewer do not trigger the rule, and GitHub documents are not publications. Separately, papers under review are embargoed from proactive media discussion until acceptance — a violation can remove the paper from the proceedings. A journalist writing about your arXiv preprint without your involvement is not a violation; you pitching the story is.
Submitting authors are enrolled into the reviewer pool. This is not symbolic: papers authored by a reviewer who blows review deadlines or files highly irresponsible reviews can be desk-rejected wholesale at PC discretion. Before the deadline, decide which coauthors will absorb December–January reviewing load and calendar it.
| Finding | Severity | Recoverable? | |---|---|---| | Coauthor without a valid OpenReview profile at abstract deadline | Desk reject | Only before Nov 6 | | Page 9 contains anything besides references | Rejected without review | Only before the paper deadline | | Style file modified / margins compressed | Rejected without review | Fix and rebuild | | Identity leak (metadata, acknowledgements, linked repo) | Desk reject | Sweep before upload | | External link expanding content | Policy violation | Delete; move content into supplement | | CRF missing | Incomplete submission | Complete Sections 1 and 5 | | Same work under review elsewhere in the window | Administrative rejection | Withdraw one venue first | | Assigned coauthor review never delivered (January) | All their papers at risk | Deliver the reviews |
The PDF gets the rehearsals; the form causes the midnight surprises. Draft these a day ahead: the abstract field (verbatim from the PDF — reviewers bid on it), subject areas (chosen for reviewer routing, not maximal coverage), conflicts (every coauthor's domain conflicts, checked against the institution list), and the author order (frozen — post-deadline author changes are a chairs-level exception, not an edit).
leaked in.
match the PDF verbatim because reviewers bid on it.
figures render at print zoom, since reviewers judge qualitative results visually.
text[CVPR submission] ready / blocked / at-risk [Profiles] all coauthors valid on OpenReview: yes/no [Format] pages used (of 8 incl. figures/tables) · refs-only overflow ok [CRF] mandatory sections complete: yes/no [Anonymity/links] clean / violations: <list> [Duty plan] reviewing coauthors named: yes/no [Fix queue] <ordered blockers>
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→pass | 16,617 | 16,588 | -0% | 1 | 1 | 0% | 2,133 | 3,614 | +69% | 0 | 0 | — |
case-02 | fail→pass | 32,811 | 19,689 | -40% | 1 | 1 | 0% | 3,687 | 4,251 | +15% | 0 | 0 | — |
case-03 | fail→fail | 36,608 | 20,818 | -43% | 1 | 1 | 0% | 4,221 | 4,277 | +1% | 0 | 0 | — |
case-04 | pass→pass | 13,214 | 13,756 | +4% | 1 | 1 | 0% | 1,402 | 3,166 | +126% | 0 | 0 | — |
case-05 | fail→pass | 17,672 | 9,900 | -44% | 1 | 1 | 0% | 1,998 | 3,401 | +70% | 0 | 0 | — |
case-06 | fail→pass | 18,632 | 12,288 | -34% | 1 | 1 | 0% | 2,125 | 3,046 | +43% | 0 | 0 | — |
case-07 | pass→pass | 17,798 | 11,592 | -35% | 1 | 1 | 0% | 1,979 | 3,132 | +58% | 0 | 0 | — |
case-08 | fail→pass | 15,404 | 13,649 | -11% | 1 | 1 | 0% | 1,670 | 3,100 | +86% | 0 | 0 | — |
case-09 | fail→pass | 16,564 | 16,414 | -1% | 1 | 1 | 0% | 1,821 | 3,184 | +75% | 0 | 0 | — |
case-10 | fail→pass | 19,531 | 13,350 | -32% | 1 | 1 | 0% | 2,373 | 3,315 | +40% | 0 | 0 | — |
case-11 | fail→pass | 11,790 | 9,616 | -18% | 1 | 1 | 0% | 1,640 | 2,437 | +49% | 0 | 0 | — |
case-12 | pass→pass | 17,728 | 10,959 | -38% | 1 | 1 | 0% | 1,739 | 2,673 | +54% | 0 | 0 | — |
case-13 | fail→fail | 17,839 | 14,196 | -20% | 1 | 1 | 0% | 1,899 | 2,854 | +50% | 0 | 0 | — |
case-14 | pass→pass | 13,296 | 12,502 | -6% | 1 | 1 | 0% | 1,319 | 2,919 | +121% | 0 | 0 | — |
case-15 | pass→pass | 22,191 | 20,694 | -7% | 1 | 1 | 0% | 3,203 | 4,508 | +41% | 0 | 0 | — |
case-16 | pass→pass | 16,573 | 14,413 | -13% | 1 | 1 | 0% | 1,889 | 3,149 | +67% | 0 | 0 | — |
case-17 | pass→pass | 15,273 | 8,894 | -42% | 1 | 1 | 0% | 1,523 | 2,828 | +86% | 0 | 0 | — |
case-18 | pass→pass | 17,718 | 8,700 | -51% | 1 | 1 | 0% | 1,901 | 2,965 | +56% | 0 | 0 | — |
case-19 | fail→pass | 25,409 | 17,342 | -32% | 1 | 1 | 0% | 3,325 | 3,917 | +18% | 0 | 0 | — |
case-20 | pass→pass | 20,664 | 13,672 | -34% | 1 | 1 | 0% | 2,288 | 3,953 | +73% | 0 | 0 | — |
case-21 | pass→pass | 16,390 | 14,208 | -13% | 1 | 1 | 0% | 2,209 | 3,970 | +80% | 0 | 0 | — |
case-22 | pass→pass | 14,441 | 17,533 | +21% | 1 | 1 | 0% | 1,433 | 3,365 | +135% | 0 | 0 | — |
case-23 | pass→fail | 9,569 | 8,198 | -14% | 1 | 1 | 0% | 1,433 | 2,955 | +106% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 23 cases were attempted. The headline lift of +35 percentage points is the difference between those two pass rates over the 23 comparable cases. 1 case got worse with the skill loaded, and it is included in that figure.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.