Paper Section Playbook

Paper Section Playbook

Scope and complement

This skill answers "how should this section be structured and what should each paragraph do?" for papers in CV, 3D perception, and autonomous driving venues (CVPR / ICCV / ECCV / NeurIPS / CoRL).

For sentence-level prose polish (word choice, punctuation, terminology consistency, em-dash rules, etc.), defer to paper-refinement-skills instead. Both skills may be active at once; this one operates at the section and paragraph level, the other at the sentence level.

When to load this skill

• The user is planning or restructuring a paper section, not just polishing prose.
• The user asks "how do I write the Intro / Method / Experiments / etc."
• The user is rebutting a reviewer comment about section structure, motivation, or framing.
• The user wants exemplar-based guidance ("how do top papers do X").

Entry mode and settled decisions

• For a new section or a requested restructuring, use the structural routing below.
• For an approved draft's result refresh, caption correction, or local clarity fix, reuse its established story and section roles. Inspect only the relevant evidence and connected passages; do not repeat elicitation or a full exemplar search.
• A new claim or changed evidence boundary warrants a focused structural review, not automatic redesign of every section. Ask only when the unresolved choice materially changes the paper's contribution or the user's requested scope.
• A missing experiment is a limitation or a draft placeholder, not authorization to train a model. Continue supported writing without inventing missing evidence.

Routing

1. Identify which section the user is working on.
2. Read the matching file under sections/ for the structural playbook (paragraph functions, decision tree, common reviewer red flags).
3. If the user is choosing between writing styles, consult decision-trees/paper-type-router.md to identify their paper type (new method / new task / new analysis / system / dataset), then apply the corresponding branch in the section file.
4. For Related Work, method motivation, experiment framing, or exemplar-based guidance, use $research-evidence when citation support or top-venue precedent is uncertain. Prefer CVPR, ICCV, ECCV, ICLR, NeurIPS/NIPS, and ICML before secondary autonomous-driving or robotics venues unless the topic requires them.
5. Use the worked skeletons and templates in the relevant section file when the user wants concrete patterns.
6. After drafting, run the section file's self-review checklist and checklists/reviewer-redflags.md for the relevant section.

Authoring principle

Section files combine patterns from cited, high-impact exemplar papers with the generalized principles in references/design-rationale.md. Treat exemplars as evidence, not authority. Keep a rule only when it has a traceable source or a clear rationale with stated conditions.

Do not invent generic writing advice. If a rule has no source, it does not belong here.

Section files

• sections/abstract.md — authored
• sections/introduction.md — authored
• sections/related-work.md — authored
• sections/method.md — authored
• sections/experiments.md — authored
• sections/conclusion.md — authored

(Files appear as the corresponding section is elicited and ratified by the user.)

Post-use feedback loop (meta-rule)

This skill is a working hypothesis, not a fixed rulebook. When a real application reveals a reusable problem, propose a scoped skill update rather than editing or logging private interactions automatically. Apply the following steps only when that update is authorized:

1. Record where a recommendation failed, drew justified pushback, or conflicted with venue or paper constraints.
2. If a rule produced a bad recommendation in context, do not silently follow it next time. Either:
   • Narrow the rule (add the condition under which it applies), or
   • Flip the rule's default (if the user's corrected choice is the better general default), or
   • Remove the rule (if it was over-generalized guidance).
3. Update references/design-rationale.md only when the lesson is generalizable; do not preserve private conversations, project identifiers, or dated interaction logs.
4. User-originating preferences always outrank inferred rules from exemplars.

Signals to watch for:

• "Why did you put X in Y?" → the placement rule is probably over-normative.
• "I don't need to follow this, I'll do Z instead." → the default may be wrong for this user's venue/style.
• "This feels defensive / redundant / apologetic." → the rule is adding ceremony, not substance.

Existing explicit approval covers its scoped skill update; do not ask again for each sentence. A materially different update or public sync still needs applicable authorization. Never publish private interaction logs or paper-specific evidence.

Cross-skill handoff

After this skill produces a section draft or restructuring plan, hand off to paper-refinement-skills for sentence-level polish. Do not duplicate prose-polish rules here. For substantial edits, also check that section-level claims still match the paper's tables, figures, citations, notation, and rendered cross-references.