A Claude Code skill repository that turns Claude into a rigorous academic research partner — from first idea to final submission.
paperstack covers the full research lifecycle:
- Idea development: Stress-test, sharpen, and frame your research ideas before committing months of work.
- Research workflows: Study a topic through structured reading, adversarial critique, cross-paper comparison, and synthesis.
- Draft review: Review your own paper before submission. Get conference-style feedback that exposes rejection risk.
Most AI paper tools produce summaries. Summaries are not understanding — and summaries won't save you from Reviewer 2.
Understanding requires you to state what you believe before reading, identify the weakest link in an argument, separate claims from evidence, and take a position you can defend. paperstack forces this discipline at every step.
Pre-submission review requires something harder: simulating the scrutiny of a hostile but fair reviewer who knows the field. paperstack's /draft-review does this by scanning your workspace, reading your full manuscript, and generating a conference-style review with verdict, major concerns, and a revision checklist.
| Without paperstack | With paperstack |
|---|---|
| "Summarize this paper" | Multi-lens reading through 6 analytical perspectives |
| Vague research questions | Forcing questions that demand precision and falsifiability |
| Cherry-picked papers | Systematic literature mapping with explicit triage criteria |
| Passive reading | Adversarial critique with claim-evidence alignment |
| "The paper is interesting" | Specific: WHAT is interesting, WHY, and what evidence supports it |
| Summary of summaries | Synthesis that identifies consensus, conflicts, and gaps |
| Submit and pray | Conference-style review before submission, with repair checklist |
| "Is my idea any good?" | Stress-test with 5 forcing questions before committing months of work |
| Skill | What it does |
|---|---|
/idea-test |
Stress-test an idea before you commit. 5 forcing questions on novelty, significance, baselines, and feasibility. Verdict: PURSUE / REFINE / PIVOT / ABANDON. |
/idea-sharpen |
Sharpen a rough idea into a precise contribution. Forces draft abstract, positioning against 3 specific papers, mock rebuttal, killer figure, and experimental plan. |
/contribution-frame |
Frame results into a publishable contribution. Forces articulation of insights over methods, tests structural fragility, preempts rejection. |
| Skill | What it does |
|---|---|
/research-intake |
Define a research question with precision. Forces honesty about priors, scope, and success criteria. |
/literature-map |
Map the research landscape — clusters, timelines, key papers, datasets, metrics, gaps. |
/paper-triage |
Prioritize papers for reading. Rank by relevance, not familiarity. |
/paper-read |
Deep-read a paper through 6 lenses: Author, Reviewer, Skeptic, Reproducer, Practitioner, Student. |
/paper-critic |
Adversarial critique — claim-evidence alignment, novelty, baseline fairness, reproducibility. |
/compare-papers |
Compare papers side-by-side on dimensions that matter for your question. |
/synthesis |
Synthesize findings into a coherent answer. No hedging. Take a position. |
| Skill | What it does |
|---|---|
/draft-review |
Review your own paper before submission. Scans your workspace, generates a conference-style review with venue-aware critique, and outputs a prioritized revision checklist. |
/idea-test Is the idea worth pursuing?
|
v
/idea-sharpen Sharpen into a precise claim (repeat until sharp)
|
v
/research-intake Define the reading question
|
v
/literature-map Map the landscape
|
v
/paper-read Deep-read each paper
|
v
/paper-critic Critique each paper
|
v
/synthesis Synthesize findings
|
v
[run experiments]
|
v
/contribution-frame Frame results into a paper
|
v
[write the draft]
|
v
/draft-review Review before submission
You don't have to run every skill. Enter at any point. The sequence is a guide, not a cage.
/idea-test
One command to stress-test an idea before committing. Get a verdict (PURSUE / REFINE / PIVOT / ABANDON), risk map, and a "next 48 hours" plan.
/draft-review
/draft-review scans your workspace for the manuscript, asks for the target venue, and generates a conference-style review with a revision checklist. Output goes to .paperstack/.
- Researchers with a new idea who want honest validation before committing months
- Authors preparing to submit who want brutally honest pre-submission feedback
- PhD students doing literature reviews or developing thesis proposals
- Engineers reading papers to understand what to implement
- Anyone stuck between results and paper who can't articulate the contribution
Open Claude Code and paste this:
Install paperstack: run
git clone https://github.com/ziyangliu-666/paperstack.git ~/.claude/skills/paperstack && ~/.claude/skills/paperstack/setupthen add a "paperstack" section to CLAUDE.md that lists the available skills: /idea-test, /idea-sharpen, /contribution-frame, /draft-review, /research-intake, /literature-map, /paper-triage, /paper-read, /paper-critic, /compare-papers, /synthesis.
Claude does the rest.
git clone https://github.com/ziyangliu-666/paperstack.git ~/.claude/skills/paperstack
~/.claude/skills/paperstack/setupThe setup script creates symlinks so Claude Code can discover each skill. Then add to your CLAUDE.md (project or ~/.claude/CLAUDE.md):
## paperstack
Available skills: /idea-test, /idea-sharpen, /contribution-frame, /draft-review,
/research-intake, /literature-map, /paper-triage, /paper-read, /paper-critic,
/compare-papers, /synthesis.Copy just the skill directories into your project:
mkdir -p .claude/skills/paperstack
for skill in ~/.claude/skills/paperstack/*/; do
[ -f "$skill/SKILL.md" ] && cp -Rf "$skill" .claude/skills/paperstack/
doneCommit .claude/skills/paperstack/ to your repo.
paperstack/
.paperstack/ # All paperstack outputs (per workspace)
ideas/ # Idea assessments from /idea-test
briefs/ # Idea briefs from /idea-sharpen
frames/ # Contribution frames from /contribution-frame
latest-review.md # Latest draft review
history/ # Timestamped review history
docs/
architecture.md # How artifact chaining works
workflows.md # Recommended sequences
research/current/ # Active research artifacts
papers/
cards/ # Paper cards (one per paper)
critiques/ # Paper critiques (one per paper)
metadata/ # Structured metadata
templates/ # Markdown templates for all artifacts
scripts/ # Optional Python helpers
examples/
sample-topic/ # Research workflow example
draft-review/ # Draft review output example
idea-test/ # Idea assessment example
idea-test/ # Skill definitions (each with SKILL.md)
idea-sharpen/
contribution-frame/
draft-review/
research-intake/
literature-map/
paper-triage/
paper-read/
paper-critic/
compare-papers/
synthesis/
setup # Install script (creates symlinks)
- Evidence first — Separate claims from interpretation. Record uncertainty. Never overstate.
- Multi-lens reading — Every paper analyzed from 6 perspectives, not just summarized.
- Artifact chaining — Each skill reads prior artifacts and writes new ones. Research builds on itself.
- Forcing questions — 33 probing questions across all workflows that demand specificity and honesty.
- Anti-sycophancy — "The paper is interesting" is banned. Say what, why, and with what evidence.
- Local-first — No database, no web service, no frontend. Plain files in your workspace.
- Loud degradation — If venue-specific guidance is unavailable, paperstack must say so explicitly and fall back to cache or generic review criteria.
Optional Python scripts (stdlib only, no dependencies):
python scripts/init_research.py "My Research Topic" # Set up directory structure
python scripts/new_paper_card.py "Lewis" "2020" "RAG" # Create a paper card
python scripts/validate_markdown.py papers/cards/*.md # Check required sections