Use when structuring the appendix of a COLT (Conference on Learning Theory) submission — the unlimited-length proof appendix inside the single PDF — covering body/appendix splitting for a 12-page limit, theorem restatement discipline, lemma ordering, notation tables, and what referees expect to find where.
SkillScore82Maintained
Automatic
◆Writing & Editing pack
Last verified 8/31/2026
What it changes
Passed 18 of 22 test cases with the skill, and 14 of 22 without it.
Without skill64%
With skill82%
Checks it passed that the model alone missed
The response identifies Appendix D as the designated section for auxiliary lemmas.
The output formats each section with bracketed line prefixes rather than markdown headings.
The response requires all theorem hypotheses to be explicitly visible in the main body statement.
and 2 more in the full report
5 fixed13 passed either way3 still failing1 broke
Pass-rate lift
+18 pp
Tokens
+34%
Agent turns
0%
Response time
-11%
Model gemini-3.6-flashMeasured 8/31/2026v118 of 22 cases passed with skill
Tokens, turns and time describe resource use, not a dollar price.
See it in action
One of 5 test cases the skill fixed. Same model, same prompt, with and without the skill.
1 of 5
Prompt
Our COLT Appendix B contains a 15-page proof of our main minimax theorem, jumping immediately into `\begin{proof} Let \epsilon > 0 ...` without any introductory explanation. Review this proof structure and output the evaluation using the five required fields.
What was checked, without and with the skill
Checked
Without skill
With skill
The response requires major appendix proofs to open with a roadmap paragraph.
Passed
Passed
The response specifies that the roadmap paragraph should explain key lemma contributions and technical novelties.
Missed
Passed
case-08 · gemini-3.6-flash · 8/31/2026 · answers are capped and redacted
What's inside
The instructions your model reads when the skill loads.
SKILL.md123 linessingle file
COLT Supplementary
COLT's "supplementary material" is the appendix of the same PDF: the 2026 CFP (checked
2026-07-08) capped the main body at 12 PMLR-formatted pages excluding references and
put no page limit on references and appendices, with the entire manuscript
submitted as a single file. There is no separate upload, no later supplement deadline,
and no distinction between "appendix" and "supplement" — which changes how you should
architect the document compared to venues with detached supplements.
Consequences of the single-PDF model
The appendix is unambiguously part of the reviewed record; referees are choosing to
read it, not to download it. Lower the cost of that choice.
Nothing can be added after the deadline. A proof hole discovered post-submission
cannot be patched by a supplement upload the way an artifact fix sometimes can
elsewhere.
Length is free but attention is not: a 60-page appendix is normal at COLT, a
disorganized one is a soft reject, because the referee's verification budget is the
binding constraint.
The 12-page split decision
What must stay in the body:
Formal problem setup: the model, the interaction protocol, the adversary or
distributional assumptions, and the performance measure — complete, not "see
Appendix A for the model."
Every main theorem statement, with all hypotheses visible.
A proof overview per main theorem: the decomposition, the key lemma, and the step
where prior techniques fail. This is what earns the referee's appendix time.
The comparison to nearest prior bounds (a small table often beats prose).
What belongs in the appendix: full proofs, auxiliary lemmas, generalized statements,
deferred case analyses, extended related work, and any numerical-illustration details.