Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Improve academic paper writing quality for ML/CV/NLP-style papers with clear section structure, paragraph flow, and reviewer-facing presentation. Use when drafting or revising Abstract, Introduction, Related Work, Method, Experiments, or Conclusion; polishing figures/tables; checking claim-support alignment; or performing self-review before submission.
.claude/skills/a-attia-research-paper-writing/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-09 | ✗→✓ | ▲ Improved | 26% | 0% |
| case-07 | ✗→✓ | ▲ Improved | 34% | 0% |
| case-10 | ✗→✓ | ▲ Improved | 61% | 0% |
| case-13 | ✗→✓ | ▲ Improved | 45% | 0% |
| case-15 | ✗→✓ | ▲ Improved | 30% | 0% |
Use this skill to rewrite a research paper into a reviewer-friendly, high-clarity draft. Prioritize first-impression quality (figures/tables/layout), logical flow, and evidence-backed claims.
references/.references/paper-review.md.Use this quick test whenever the user asks whether a paragraph "flows" or is clear.
Source reference for this check:
references/does-my-writing-flow-source.mdLoad only the needed section file:
references/introduction.mdreferences/abstract.mdreferences/related-work.mdreferences/method.mdreferences/experiments.mdreferences/conclusion.mdreferences/paper-review.mdreferences/does-my-writing-flow-source.mdreferences/examples/index.mdUse references/paper-review.md for the full checklist and workflow.
When asked to rewrite or draft sections, return:
Claim: ... | Evidence: ... | Status: supported/needs evidence.This skill governs the published paper draft itself (Abstract, Introduction, Related Work, Method, Experiments, Conclusion). When the work also requires producing or revising meta-documents that support the paper -- for example:
PLAN.md,notes/section_<N>.md),notes/impl_<component>.md),README.md,those documents have a different audience (the user, co-authors, reviewers) and a different style budget than the paper itself. Load ~/.scicomp-research-skills/skills/human-facing-doc-authoring/SKILL.md for the conventions and per-doc-type structural skeletons. Use BOTH skills together: paper-writing for the paper draft, human-facing for any meta-doc.
A paper-writing session typically iterates across multiple sections, re-reads many survey notes, may re-grep the paper draft repeatedly, and often spans multiple sessions before a section is finalised. Whenever the session is going to be heavy (more than ~5 file reads, multiple section drafts, or a span across multiple sittings), load ~/.scicomp-research-skills/skills/agent-resource-discipline/SKILL.md at session start. In particular:
references/persistent-memory.md -- the first-/last-actionprotocol that updates notes/section_<N>.md and PLAN.md status fields so the next session resumes cheaply.
references/context-window-budget.md -- the rule that this skill'sExecution Rule #7 ("do not load all section references at once") generalises to: do not load all survey notes at once, do not load the full .txt extraction of every cited paper -- target-grep instead.
Both skills compose: paper-writing for what to write; agent-resource- discipline for what to load while writing.
The references files under references/ (abstract.md, conclusion.md, experiments.md, introduction.md, method.md, paper-review.md, related-work.md, examples/) are vendored from the upstream Master-cai/Research-Paper-Writing-Skills repository at commit 9ee5edd (see ATTRIBUTION.md). They are deliberately preserved without per-file date-stamp footers -- our universal "every plan-of-record-style document ends with a date-stamp footer" convention (root AGENTS.md Section 6) does not apply to upstream-imported reference content; modifying the upstream files to add footers would create false revision history.
When upstream evolves the references files, we may selectively cherry-pick changes (per ATTRIBUTION.md "Pulling updates from upstream"). When we substantially modify a reference file, we will add a *Revised YYYY-MM-DD by A. Attia: <note>.* footer to that specific file at that point. Until then the upstream provenance is authoritative.
Revised 2026-05-13: added "Adjacent human-facing documents" section pointing at the human-facing-doc-authoring skill for meta-docs (PLAN.md, section/impl notes, README, rebuttal) that this skill does not directly govern. Revised 2026-05-13: added "Resource discipline for paper-writing sessions" section pointing at the agent-resource-discipline skill for the cross-cutting context-window / persistent-memory protocol that complements this skill's Execution Rule #7. Revised 2026-05-14 (post-fresh-audit): documented the deliberate exception from the universal date-stamp-footer convention for upstream-vendored references files (the eight files under references/ that were imported from Master-cai upstream and have not been substantially modified).
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 15,466 | 4,098 | -74% | 1 | 1 | 0% | 3,074 | 2,259 | -27% | 0 | 0 | — |
case-02 | fail→fail | 14,126 | 4,113 | -71% | 1 | 1 | 0% | 3,220 | 2,259 | -30% | 0 | 0 | — |
case-03 | fail→fail | 26,333 | 15,817 | -40% | 1 | 1 | 0% | 4,165 | 4,441 | +7% | 0 | 0 | — |
case-09 | fail→pass | 11,101 | 3,431 | -69% | 1 | 1 | 0% | 2,087 | 2,622 | +26% | 0 | 0 | — |
case-04 | fail→fail | 15,059 | 14,682 | -3% | 1 | 1 | 0% | 3,046 | 4,170 | +37% | 0 | 0 | — |
case-05 | fail→fail | 14,571 | 13,150 | -10% | 1 | 1 | 0% | 2,737 | 3,916 | +43% | 0 | 0 | — |
case-06 | pass→fail | 8,886 | 5,589 | -37% | 1 | 1 | 0% | 1,753 | 2,439 | +39% | 0 | 0 | — |
case-07 | fail→pass | 9,805 | 3,012 | -69% | 1 | 1 | 0% | 1,870 | 2,513 | +34% | 0 | 0 | — |
case-08 | pass→pass | 7,862 | 5,221 | -34% | 1 | 1 | 0% | 1,564 | 2,937 | +88% | 0 | 0 | — |
case-10 | fail→pass | 8,630 | 3,086 | -64% | 1 | 1 | 0% | 1,591 | 2,568 | +61% | 0 | 0 | — |
case-11 | pass→pass | 8,645 | 6,961 | -19% | 1 | 1 | 0% | 1,748 | 3,268 | +87% | 0 | 0 | — |
case-12 | pass→pass | 7,866 | 5,152 | -35% | 1 | 1 | 0% | 1,473 | 2,893 | +96% | 0 | 0 | — |
case-13 | fail→pass | 7,956 | 1,634 | -79% | 1 | 1 | 0% | 1,536 | 2,225 | +45% | 0 | 0 | — |
case-14 | pass→pass | 13,345 | 6,410 | -52% | 1 | 1 | 0% | 2,834 | 2,827 | -0% | 0 | 0 | — |
case-15 | fail→pass | 11,800 | 4,277 | -64% | 1 | 1 | 0% | 2,203 | 2,863 | +30% | 0 | 0 | — |
case-16 | fail→fail | 11,358 | 5,224 | -54% | 1 | 1 | 0% | 2,148 | 2,943 | +37% | 0 | 0 | — |
case-17 | pass→pass | 10,492 | 4,212 | -60% | 1 | 1 | 0% | 1,973 | 2,631 | +33% | 0 | 0 | — |
case-18 | fail→fail | 10,712 | 3,908 | -64% | 1 | 1 | 0% | 2,015 | 2,200 | +9% | 0 | 0 | — |
case-19 | pass→fail | 10,481 | 4,927 | -53% | 1 | 1 | 0% | 1,869 | 2,294 | +23% | 0 | 0 | — |
case-20 | pass→pass | 10,382 | 8,405 | -19% | 1 | 1 | 0% | 2,832 | 4,097 | +45% | 0 | 0 | — |
case-21 | pass→pass | 12,765 | 14,754 | +16% | 1 | 1 | 0% | 2,750 | 5,027 | +83% | 0 | 0 | — |
case-22 | pass→pass | 12,209 | 15,916 | +30% | 1 | 1 | 0% | 2,217 | 5,124 | +131% | 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. 22 cases were attempted, and 16 counted toward the lift figure. The other 6 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 +14 percentage points is the difference between those two pass rates over the 16 comparable cases. 3 cases got worse with the skill loaded, and they are 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.