Deep Review
dipakkrishnan/deep-review-skill
Revisão crítica profunda de papers e artefatos técnicos com calibração e memória
Ver no GitHubSkill para revisão rigorosa e crítica adversarial de artigos e artefatos técnicos. Realiza entrevista de calibração, decompõe a revisão em passes especializadas, valida afirmações contra fontes e computações, e retorna feedback estruturado por severidade. Ideal para pré-submissão de papers, testes de provas e auditorias metodológicas.
Instalação
npx convez-ai dipakkrishnan/deep-review-skill1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Revisão de artigos antes de submissão
- Verificação de correção de teoremas e provas matemáticas
- Auditoria de métodos e experimentos empíricos
- Análise de posicionamento em literatura científica
- Crítica de rascunhos técnicos e gerados por IA
Veredito

Deep Review é uma ferramenta niche mas poderosa para pesquisadores e revisores que precisam de critique automática rigorosa. Está em estágio experimental com distribuição pessoal do autor, recomendado para quem tem workflows críticos de revisão técnica.
README
A tradução para o português deste README ainda está na fila.
Deep Review
Deep Review is an experimental skill for rigorous review of papers and technical artifacts.
It asks a short calibration interview, reads the artifact carefully, decomposes the review into specialist passes, checks claims against sources and computations where possible, self-critiques its own findings, and returns a severity-ranked review.
The goal is simple: find the errors a generic chat-with-file pass is likely to miss.
Deep Review can also maintain optional local review memory in .deep-review/, so future reviews can pick up the user's preferred standards, recurring concerns, and review taste. It only writes memory after explicit user approval.
What It Is Good For
- pre-submission paper review
- theorem or proof stress tests
- empirical / methods audits
- literature positioning checks
- technical memo review
- AI-generated draft review
Status
This is early. I am distributing it personally for now while I learn where it is most useful.
If you want to try it on a real paper or artifact, open an issue or reach out.
Install
Claude Code
Add the direct Deep Review marketplace:
/plugin marketplace add dipakkrishnan/deep-review-skill
/plugin install deep-review@deep-review-marketplace
/reload-plugins
This is the fastest install path while the approved community marketplace listing propagates.
Or copy the skill folder into a project directly:
mkdir -p .claude/skills
cp -R /path/to/deep-review-skill/skills/deep-review .claude/skills/deep-review
Then ask Claude Code:
Use $deep-review to review this paper for theorem correctness and literature positioning.
Claude.ai
Download deep-review.zip (always points at the latest release), then in Claude.ai go to Customize → Skills → + → Upload a skill and drop the zip in.
To build the zip yourself from a checkout:
scripts/package-skill.sh
# produces dist/deep-review.zip
Codex
Codex does not yet expose a marketplace-style installer for third-party plugins. Copy the skill into a Codex skills directory and restart Codex:
cp -R skills/deep-review "$CODEX_HOME/skills/deep-review"
Example Prompts
Use $deep-review to review this PDF as a skeptical economics referee. Focus on model correctness, calibration, and missing literature.
Use $deep-review to stress-test the proof of Theorem 2 and identify assumptions that do not hold.
Use $deep-review to review this technical memo. Skip copy edits and focus on unsupported claims.
What To Expect
Deep Review will usually ask a few calibration questions first:
- What is your relationship to the artifact?
- What kind of criticism is most useful?
- Which parts worry you most?
- What should it avoid spending time on?
Then it produces a markdown review with:
- overall assessment
- critical issues
- significant concerns
- minor issues
- questions for the authors
Optionally, Deep Review can save local memory for future runs:
- standing review preferences
- target venues or standards
- recurring issue patterns
- lightweight history of reviewed artifacts
Requirements
The skill itself is instructions and references.
The helper scripts require Python. PDF extraction uses PyMuPDF:
pip install pymupdf
Privacy
Deep Review does not collect telemetry or send data to any third-party service controlled by this plugin. See PRIVACY.md.
Notes
Deep Review is not a replacement for expert human judgment. It is a way to make review cheaper, sharper, and more adversarial before you ask humans to spend scarce attention on the artifact.