Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Use when preparing or auditing a CoRL OpenReview submission — the corl_2026 LaTeX template, the 8-page main text with a mandatory Limitations section counted inside it, uncounted references and appendix, the supplementary file and 250 MB video, double-anonymous rules, abstract registration, and dual-submission checks.
.claude/skills/brycewang-stanford-corl-submission/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-02 | ✗→✓ | ▲ Improved | 29% | 0% |
| case-04 | ✗→✓ | ▲ Improved | 72% | 0% |
| case-08 | ✗→✓ | ▲ Improved | 121% | 0% |
| case-09 | ✗→✓ | ▲ Improved | 44% | 0% |
| case-15 | ✗→✓ | ▲ Improved | 68% | 0% |
This is the pre-upload audit for a CoRL paper. The numbers below are the CoRL 2026 requirements as read on 2026-07-08 from the official Instruction for Authors and Call for Papers (corl.org); the 2026 deadline has passed, so apply this skill to a 2027-cycle draft after re-opening the new year's pages — CoRL re-issues its instructions annually and has changed page accounting between cycles.
| Element | 2026 rule | Notes | |---|---|---| | Platform | OpenReview, group robot-learning.org/CoRL/2026/Conference | Profile + conflicts needed for every author | | Abstract registration | May 26, 2026 | Separate, earlier than the paper | | Paper deadline | May 29, 2026 | Deadline clock/timezone wording 待核实 per cycle | | Main text | 8 pages, including a mandatory Limitations section | Limitations counts toward the limit — budget for it | | Not counted | Acknowledgments, references, optional appendix | Appendix rides in the same PDF | | Template | corl_2026 LaTeX class | Camera-ready later flips to [final] mode | | Review model | Double-anonymous | Applies to paper, supplementary, and any linked site | | Supplementary | Encouraged; video ≤ 250 MB (strict), ~3 min suggested | See corl-supplementary for content strategy |
Two venue-specific traps dominate desk problems here, and neither exists in the same form at neighboring conferences:
require every submission to include a Limitations section — limiting assumptions, failure modes, and how they might be addressed — and it counts inside the 8 pages. Drafts imported from NeurIPS-style templates (where such sections often sit outside the limit) silently overflow. Audit the page budget with Limitations present, not as an afterthought.
project website with videos. During review, any link that identifies an author, lab, or institution is an anonymity violation under the 2026 instructions — this includes the URL itself (lab-name.github.io fails even if the page is scrubbed). Use an anonymous hosting pattern or put the video in the supplementary upload instead.
Beyond the standard items (author block, acknowledgments, metadata, third-person self-citation), sweep the modalities specific to this field:
robot shells, distinctive lab furniture that appears in your prior papers.
associated with your group.
/home/<user>/...).bash# Mechanical pre-upload pass on the submission PDF and supplementary bundle pdftotext main.pdf - | grep -nEi 'github\.io|github\.com|wandb|acknowledg|our (lab|group|university)' | head pdfinfo main.pdf | grep -Ei 'author|creator|producer' grep -RnEi 'hostname|/home/|api_key' supplementary/code/ | head stat -c '%n %s' supplementary/video.mp4 # must be <= 250 MB (2026 strict cap) pdftotext main.pdf - | grep -ci 'limitation' # 0 hits => missing mandatory section
a common shape); do not let it become the compression victim of deadline week.
non-load-bearing detail (full hyperparameter tables, extra rollouts, per-task breakdowns beyond the headline) may live there.
float spacing are detectable and read as rule-testing to ACs.
paper (May 29); a missed abstract registration forecloses the submission entirely. Register with a stable title and author list.
fast-moving robot-learning arXiv culture makes this an easy accident when a workshop version is circulating — check the current CFP's exact scope wording each cycle.
(announced April 2026); its exact disclosure requirements are 待核实 in this pack — read that section verbatim before submitting.
specialized pool; pick the subfield your harshest competent critic works in.
text[ ] Abstract registered by the abstract deadline (title/authors stable) [ ] Main text <= 8 pages WITH Limitations section included and substantive [ ] References + appendix after main text, same PDF, uncounted [ ] corl_2026 template unmodified; compiles with no page-limit tricks [ ] Anonymity sweep passed: text, figures, video frames, links, metadata [ ] Supplementary video <= 250 MB, ~3 min, anonymized end to end [ ] Code/data (if included) scrubbed of identities, paths, and keys [ ] OpenReview form: authors, conflicts, keywords, abstract synced with PDF [ ] Dual-submission status confirmed with every coauthor in writing [ ] GenAI-policy disclosure handled per the current instructions [待核实]
Page accounting (the 2026 camera-ready got a 9th page; earlier cycles differed), the Limitations rule, video caps, the GenAI policy, and deadline clock conventions are all chair-set and cycle-volatile. The controlling sources are https://www.corl.org/contributions/call-for-papers and https://www.corl.org/contributions/instruction-for-authors for the live year — open both before trusting any number in this file.
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 50,374 | 30,191 | -40% | 1 | 1 | 0% | 7,743 | 4,856 | -37% | 0 | 0 | — |
case-02 | fail→pass | 30,080 | 25,306 | -16% | 1 | 1 | 0% | 3,780 | 4,860 | +29% | 0 | 0 | — |
case-03 | fail→fail | 32,820 | 22,456 | -32% | 1 | 1 | 0% | 4,332 | 5,151 | +19% | 0 | 0 | — |
case-04 | fail→pass | 21,524 | 9,801 | -54% | 1 | 1 | 0% | 1,338 | 2,304 | +72% | 0 | 0 | — |
case-05 | pass→pass | 11,672 | 12,779 | +9% | 1 | 1 | 0% | 1,875 | 2,780 | +48% | 0 | 0 | — |
case-06 | pass→pass | 16,250 | 11,734 | -28% | 1 | 1 | 0% | 1,759 | 2,587 | +47% | 0 | 0 | — |
case-07 | pass→pass | 20,724 | 19,239 | -7% | 1 | 1 | 0% | 2,283 | 3,678 | +61% | 0 | 0 | — |
case-08 | fail→pass | 45,091 | 12,481 | -72% | 1 | 1 | 0% | 1,252 | 2,761 | +121% | 0 | 0 | — |
case-09 | fail→pass | 17,082 | 13,671 | -20% | 1 | 1 | 0% | 2,010 | 2,898 | +44% | 0 | 0 | — |
case-10 | pass→pass | 15,302 | 4,365 | -71% | 1 | 1 | 0% | 1,816 | 2,197 | +21% | 0 | 0 | — |
case-11 | pass→pass | 10,027 | 10,700 | +7% | 1 | 1 | 0% | 1,671 | 2,472 | +48% | 0 | 0 | — |
case-12 | pass→pass | 12,487 | 8,578 | -31% | 1 | 1 | 0% | 1,626 | 2,125 | +31% | 0 | 0 | — |
case-13 | pass→pass | 16,922 | 16,124 | -5% | 1 | 1 | 0% | 1,692 | 3,315 | +96% | 0 | 0 | — |
case-14 | fail→fail | 13,084 | 13,524 | +3% | 1 | 1 | 0% | 1,351 | 2,631 | +95% | 0 | 0 | — |
case-15 | fail→pass | 11,936 | 6,900 | -42% | 1 | 1 | 0% | 1,052 | 1,769 | +68% | 0 | 0 | — |
case-16 | pass→pass | 8,773 | 7,461 | -15% | 1 | 1 | 0% | 669 | 1,911 | +186% | 0 | 0 | — |
case-17 | pass→pass | 20,215 | 27,834 | +38% | 1 | 1 | 0% | 2,170 | 4,254 | +96% | 0 | 0 | — |
case-18 | pass→pass | 17,285 | 14,453 | -16% | 1 | 1 | 0% | 1,566 | 2,743 | +75% | 0 | 0 | — |
case-19 | pass→pass | 20,084 | 11,021 | -45% | 1 | 1 | 0% | 1,960 | 3,065 | +56% | 0 | 0 | — |
case-20 | pass→pass | 15,728 | 12,656 | -20% | 1 | 1 | 0% | 1,453 | 2,762 | +90% | 0 | 0 | — |
case-21 | fail→pass | 12,287 | 18,484 | +50% | 1 | 1 | 0% | 996 | 3,199 | +221% | 0 | 0 | — |
case-22 | fail→fail | 13,010 | 11,478 | -12% | 1 | 1 | 0% | 1,228 | 3,066 | +150% | 0 | 0 | — |
case-23 | pass→pass | 32,608 | 46,153 | +42% | 1 | 1 | 0% | 3,240 | 5,373 | +66% | 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, and 22 counted toward the lift figure. The other 1 produced results that are not comparable between the two arms, so they are excluded from the headline rather than averaged into it. The headline lift of +26 percentage points is the difference between those two pass rates over the 22 comparable cases.
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.